initial commit

SmartCrop/CMakeLists.txt

@@ -0,0 +1,16 @@
+cmake_minimum_required(VERSION 3.6)
+
+find_package(OpenCV REQUIRED)
+
+set(CMAKE_CXX_STANDARD 17)
+
+set(SRC_FILES main.cpp yolo.cpp tokenize.cpp log.cpp seamcarving.cpp utils.cpp intelligentroi.cpp facerecognizer.cpp)
+
+add_executable(smartcrop ${SRC_FILES})
+target_link_libraries(smartcrop ${OpenCV_LIBS} -ltbb)
+target_include_directories(smartcrop PRIVATE ${OpenCV_INCLUDE_DIRS})
+target_compile_options(smartcrop PRIVATE -s -g -Wall)
+message(WARNING ${WEIGHT_DIR})
+target_compile_definitions(smartcrop PUBLIC WEIGHT_DIR="${WEIGHT_DIR}")
+
+install(TARGETS smartcrop RUNTIME DESTINATION bin)

SmartCrop/README.md

@@ -0,0 +1,50 @@
+# SmartCrop
+
+SmartCrop is an application that uses content aware croping using, [seam carving](https://en.wikipedia.org/wiki/Seam_carving) and resizeing to bring a directory of images into the deisred size and aspect ratio for training. SmartCrop ist configurable to prioritize specific items or specifc persons in the images provided.
+
+## Requirements
+
+* [cmake](https://cmake.org/) 3.6 or later
+* [opencv](https://opencv.org/) 4.8 or later
+* A c++17 capable compiler and standard lib like gcc or llvm/clang
+* git is required to get the source
+
+## Building
+
+The steps to build this application are:
+
+	$ git clone https://uvos.xyz/git/uvos/SDImagePreprocess.git
+	$ cd SDImagePreprocess
+	$ mkdir build
+	$ cmake ..
+	$ make
+
+The binary can then be found in build/SmartCrop and can optionaly be installed with:
+
+	$ sudo make install
+
+## Basic usage
+
+To process all images in the directory ~/images and output the images into ~/proceesedImages:
+
+	$ smartcrop --out processedImages ~/images/*
+
+To also focus on the person in the image ~/person.jpg
+
+	$ smartcrop --out processedImages --focus-person ~/person.jpg ~/images/*
+
+To also enable seam carving
+
+	$ smartcrop --out processedImages --focus-person ~/person.jpg --seam-carving ~/images/*
+
+see smartcrop --help for more
+
+## Example
+
+#### Content detected in image:
+![Content found in image](images/IMGP3692.jpg)
+
+#### Cropped image based on content:
+![Cropped image](images/IMGP3692C.jpg)
+
+

SmartCrop/facerecognizer.cpp

@@ -0,0 +1,163 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include "facerecognizer.h"
+#include <filesystem>
+
+#define INCBIN_PREFIX r
+#include "incbin.h"
+
+INCBIN(defaultRecognizer, WEIGHT_DIR "/face_recognition_sface_2021dec.onnx");
+INCBIN(defaultDetector, WEIGHT_DIR "/face_detection_yunet_2023mar.onnx");
+
+#include <opencv2/dnn/dnn.hpp>
+#include <opencv2/core.hpp>
+#include <opencv2/highgui.hpp>
+#include <fstream>
+
+#include "log.h"
+
+static const std::vector<unsigned char> onnx((unsigned char*)rdefaultDetectorData, ((unsigned char*)rdefaultDetectorData)+rdefaultDetectorSize);
+
+FaceRecognizer::FaceRecognizer(std::filesystem::path recognizerPath, const std::filesystem::path& detectorPath, const std::vector<cv::Mat>& referances)
+{
+	if(detectorPath.empty())
+	{
+		Log(Log::INFO)<<"Using builtin face detection model";
+
+		detector = cv::FaceDetectorYN::create("onnx", onnx, std::vector<unsigned char>(), {320, 320}, 0.6, 0.3, 5000, cv::dnn::Backend::DNN_BACKEND_OPENCV, cv::dnn::Target::DNN_TARGET_CPU);
+		if(!detector)
+			throw LoadException("Unable to load detector network from built in file");
+	}
+	else
+	{
+		detector = cv::FaceDetectorYN::create(detectorPath, "", {320, 320}, 0.6, 0.3, 5000, cv::dnn::Backend::DNN_BACKEND_OPENCV, cv::dnn::Target::DNN_TARGET_CPU);
+		if(!detector)
+			throw LoadException("Unable to load detector network from "+detectorPath.string());
+	}
+
+	bool defaultNetwork = recognizerPath.empty();
+
+	if(defaultNetwork)
+	{
+		Log(Log::INFO)<<"Using builtin face recognition model";
+		recognizerPath = cv::tempfile("onnx");
+		std::ofstream file(recognizerPath);
+		if(!file.is_open())
+			throw LoadException("Unable open temporary file at "+recognizerPath.string());
+		Log(Log::DEBUG)<<"Using "<<recognizerPath<<" as temporary file for onnx recongnition network";
+		file.write(reinterpret_cast<const char*>(rdefaultRecognizerData), rdefaultRecognizerSize);
+		file.close();
+	}
+
+	recognizer = cv::FaceRecognizerSF::create(recognizerPath.string(), "", cv::dnn::Backend::DNN_BACKEND_OPENCV, cv::dnn::Target::DNN_TARGET_CPU);
+
+	if(defaultNetwork)
+		std::filesystem::remove(recognizerPath);
+
+	if(!recognizer)
+		throw LoadException("Unable to load recognizer network from "+recognizerPath.string());
+
+	addReferances(referances);
+}
+
+cv::Mat FaceRecognizer::detectFaces(const cv::Mat& input)
+{
+	detector->setInputSize(input.size());
+	cv::Mat faces;
+	detector->detect(input, faces);
+	return faces;
+}
+
+bool FaceRecognizer::addReferances(const std::vector<cv::Mat>& referances)
+{
+	bool ret = false;
+	for(const cv::Mat& image : referances)
+	{
+		cv::Mat faces = detectFaces(image);
+		assert(faces.cols == 15);
+		if(faces.empty())
+		{
+			Log(Log::WARN)<<"A referance image provided dose not contian any face";
+			continue;
+		}
+		if(faces.rows > 1)
+			Log(Log::WARN)<<"A referance image provided contains more than one face, only the first detected face will be considered";
+		cv::Mat cropedImage;
+		recognizer->alignCrop(image, faces.row(0), cropedImage);
+		cv::Mat features;
+		recognizer->feature(cropedImage, features);
+		referanceFeatures.push_back(features.clone());
+		ret = true;
+	}
+
+	return ret;
+}
+
+void FaceRecognizer::setThreshold(double threasholdIn)
+{
+	threshold = threasholdIn;
+}
+
+double FaceRecognizer::getThreshold()
+{
+	return threshold;
+}
+
+void FaceRecognizer::clearReferances()
+{
+	referanceFeatures.clear();
+}
+
+FaceRecognizer::Detection FaceRecognizer::isMatch(const cv::Mat& input, bool alone)
+{
+	cv::Mat faces = detectFaces(input);
+
+	Detection bestMatch;
+	bestMatch.confidence = 0;
+	bestMatch.person = -1;
+
+	if(alone && faces.rows > 1)
+	{
+		bestMatch.person = -2;
+		return bestMatch;
+	}
+
+	for(int i = 0; i < faces.rows; ++i)
+	{
+		cv::Mat face;
+		recognizer->alignCrop(input, faces.row(i), face);
+		cv::Mat features;
+		recognizer->feature(face, features);
+		features = features.clone();
+		for(size_t referanceIndex = 0; referanceIndex < referanceFeatures.size(); ++referanceIndex)
+		{
+			double score = recognizer->match(referanceFeatures[referanceIndex], features, cv::FaceRecognizerSF::FR_COSINE);
+			if(score > threshold && score > bestMatch.confidence)
+			{
+				bestMatch.confidence = score;
+				bestMatch.person = referanceIndex;
+				bestMatch.rect = cv::Rect(faces.at<int>(i, 0), faces.at<int>(i, 1), faces.at<int>(i, 2), faces.at<int>(i, 3));
+			}
+		}
+	}
+
+	return bestMatch;
+}

SmartCrop/facerecognizer.h

@@ -0,0 +1,67 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+#include <exception>
+#include <opencv2/core/mat.hpp>
+#include <opencv2/objdetect/face.hpp>
+#include <opencv2/core.hpp>
+#include <vector>
+#include <memory>
+#include <filesystem>
+
+class FaceRecognizer
+{
+public:
+
+	struct Detection
+	{
+		int person;
+		float confidence;
+		cv::Rect rect;
+	};
+
+	class LoadException : public std::exception
+	{
+	private:
+		std::string message;
+	public:
+		LoadException(const std::string& msg): std::exception(), message(msg) {}
+		virtual const char* what() const throw() override
+		{
+			return message.c_str();
+		}
+	};
+
+private:
+	std::vector<cv::Mat> referanceFeatures;
+	std::shared_ptr<cv::FaceRecognizerSF> recognizer;
+	std::shared_ptr<cv::FaceDetectorYN> detector;
+
+	double threshold = 0.363;
+
+public:
+	FaceRecognizer(std::filesystem::path recognizerPath = "", const std::filesystem::path& detectorPath = "", const std::vector<cv::Mat>& referances = std::vector<cv::Mat>());
+	cv::Mat detectFaces(const cv::Mat& input);
+	Detection isMatch(const cv::Mat& input, bool alone = false);
+	bool addReferances(const std::vector<cv::Mat>& referances);
+	void setThreshold(double threashold);
+	double getThreshold();
+	void clearReferances();
+};

SmartCrop/incbin.h

@@ -0,0 +1,495 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+/**
+ * @file incbin.h
+ * @author Dale Weiler
+ * @brief Utility for including binary files
+ *
+ * Facilities for including binary files into the current translation unit and
+ * making use from them externally in other translation units.
+ */
+#ifndef INCBIN_HDR
+#define INCBIN_HDR
+#include <limits.h>
+#if   defined(__AVX512BW__) || \
+      defined(__AVX512CD__) || \
+      defined(__AVX512DQ__) || \
+      defined(__AVX512ER__) || \
+      defined(__AVX512PF__) || \
+      defined(__AVX512VL__) || \
+      defined(__AVX512F__)
+# define INCBIN_ALIGNMENT_INDEX 6
+#elif defined(__AVX__)      || \
+      defined(__AVX2__)
+# define INCBIN_ALIGNMENT_INDEX 5
+#elif defined(__SSE__)      || \
+      defined(__SSE2__)     || \
+      defined(__SSE3__)     || \
+      defined(__SSSE3__)    || \
+      defined(__SSE4_1__)   || \
+      defined(__SSE4_2__)   || \
+      defined(__neon__)     || \
+      defined(__ARM_NEON)   || \
+      defined(__ALTIVEC__)
+# define INCBIN_ALIGNMENT_INDEX 4
+#elif ULONG_MAX != 0xffffffffu
+# define INCBIN_ALIGNMENT_INDEX 3
+# else
+# define INCBIN_ALIGNMENT_INDEX 2
+#endif
+
+/* Lookup table of (1 << n) where `n' is `INCBIN_ALIGNMENT_INDEX' */
+#define INCBIN_ALIGN_SHIFT_0 1
+#define INCBIN_ALIGN_SHIFT_1 2
+#define INCBIN_ALIGN_SHIFT_2 4
+#define INCBIN_ALIGN_SHIFT_3 8
+#define INCBIN_ALIGN_SHIFT_4 16
+#define INCBIN_ALIGN_SHIFT_5 32
+#define INCBIN_ALIGN_SHIFT_6 64
+
+/* Actual alignment value */
+#define INCBIN_ALIGNMENT \
+    INCBIN_CONCATENATE( \
+        INCBIN_CONCATENATE(INCBIN_ALIGN_SHIFT, _), \
+        INCBIN_ALIGNMENT_INDEX)
+
+/* Stringize */
+#define INCBIN_STR(X) \
+    #X
+#define INCBIN_STRINGIZE(X) \
+    INCBIN_STR(X)
+/* Concatenate */
+#define INCBIN_CAT(X, Y) \
+    X ## Y
+#define INCBIN_CONCATENATE(X, Y) \
+    INCBIN_CAT(X, Y)
+/* Deferred macro expansion */
+#define INCBIN_EVAL(X) \
+    X
+#define INCBIN_INVOKE(N, ...) \
+    INCBIN_EVAL(N(__VA_ARGS__))
+/* Variable argument count for overloading by arity */
+#define INCBIN_VA_ARG_COUNTER(_1, _2, _3, N, ...) N
+#define INCBIN_VA_ARGC(...) INCBIN_VA_ARG_COUNTER(__VA_ARGS__, 3, 2, 1, 0)
+
+/* Green Hills uses a different directive for including binary data */
+#if defined(__ghs__)
+#  if (__ghs_asm == 2)
+#    define INCBIN_MACRO ".file"
+/* Or consider the ".myrawdata" entry in the ld file */
+#  else
+#    define INCBIN_MACRO "\tINCBIN"
+#  endif
+#else
+#  define INCBIN_MACRO ".incbin"
+#endif
+
+#ifndef _MSC_VER
+#  define INCBIN_ALIGN \
+    __attribute__((aligned(INCBIN_ALIGNMENT)))
+#else
+#  define INCBIN_ALIGN __declspec(align(INCBIN_ALIGNMENT))
+#endif
+
+#if defined(__arm__) || /* GNU C and RealView */ \
+    defined(__arm) || /* Diab */ \
+    defined(_ARM) /* ImageCraft */
+#  define INCBIN_ARM
+#endif
+
+#ifdef __GNUC__
+/* Utilize .balign where supported */
+#  define INCBIN_ALIGN_HOST ".balign " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n"
+#  define INCBIN_ALIGN_BYTE ".balign 1\n"
+#elif defined(INCBIN_ARM)
+/*
+ * On arm assemblers, the alignment value is calculated as (1 << n) where `n' is
+ * the shift count. This is the value passed to `.align'
+ */
+#  define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT_INDEX) "\n"
+#  define INCBIN_ALIGN_BYTE ".align 0\n"
+#else
+/* We assume other inline assembler's treat `.align' as `.balign' */
+#  define INCBIN_ALIGN_HOST ".align " INCBIN_STRINGIZE(INCBIN_ALIGNMENT) "\n"
+#  define INCBIN_ALIGN_BYTE ".align 1\n"
+#endif
+
+/* INCBIN_CONST is used by incbin.c generated files */
+#if defined(__cplusplus)
+#  define INCBIN_EXTERNAL extern "C"
+#  define INCBIN_CONST    extern const
+#else
+#  define INCBIN_EXTERNAL extern
+#  define INCBIN_CONST    const
+#endif
+
+/**
+ * @brief Optionally override the linker section into which size and data is
+ * emitted.
+ * 
+ * @warning If you use this facility, you might have to deal with
+ * platform-specific linker output section naming on your own.
+ */
+#if !defined(INCBIN_OUTPUT_SECTION)
+#  if defined(__APPLE__)
+#    define INCBIN_OUTPUT_SECTION ".const_data"
+#  else
+#    define INCBIN_OUTPUT_SECTION ".rodata"
+#  endif
+#endif
+
+/**
+ * @brief Optionally override the linker section into which data is emitted.
+ *
+ * @warning If you use this facility, you might have to deal with
+ * platform-specific linker output section naming on your own.
+ */
+#if !defined(INCBIN_OUTPUT_DATA_SECTION)
+#  define INCBIN_OUTPUT_DATA_SECTION INCBIN_OUTPUT_SECTION
+#endif
+
+/**
+ * @brief Optionally override the linker section into which size is emitted.
+ *
+ * @warning If you use this facility, you might have to deal with
+ * platform-specific linker output section naming on your own.
+ * 
+ * @note This is useful for Harvard architectures where program memory cannot
+ * be directly read from the program without special instructions. With this you
+ * can chose to put the size variable in RAM rather than ROM.
+ */
+#if !defined(INCBIN_OUTPUT_SIZE_SECTION)
+#  define INCBIN_OUTPUT_SIZE_SECTION INCBIN_OUTPUT_SECTION
+#endif
+
+#if defined(__APPLE__)
+#  include "TargetConditionals.h"
+#  if defined(TARGET_OS_IPHONE) && !defined(INCBIN_SILENCE_BITCODE_WARNING)
+#    warning "incbin is incompatible with bitcode. Using the library will break upload to App Store if you have bitcode enabled. Add `#define INCBIN_SILENCE_BITCODE_WARNING` before including this header to silence this warning."
+#  endif
+/* The directives are different for Apple branded compilers */
+#  define INCBIN_SECTION         INCBIN_OUTPUT_SECTION "\n"
+#  define INCBIN_GLOBAL(NAME)    ".globl " INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n"
+#  define INCBIN_INT             ".long "
+#  define INCBIN_MANGLE          "_"
+#  define INCBIN_BYTE            ".byte "
+#  define INCBIN_TYPE(...)
+#else
+#  define INCBIN_SECTION         ".section " INCBIN_OUTPUT_SECTION "\n"
+#  define INCBIN_GLOBAL(NAME)    ".global " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME "\n"
+#  if defined(__ghs__)
+#    define INCBIN_INT           ".word "
+#  else
+#    define INCBIN_INT           ".int "
+#  endif
+#  if defined(__USER_LABEL_PREFIX__)
+#    define INCBIN_MANGLE        INCBIN_STRINGIZE(__USER_LABEL_PREFIX__)
+#  else
+#    define INCBIN_MANGLE        ""
+#  endif
+#  if defined(INCBIN_ARM)
+/* On arm assemblers, `@' is used as a line comment token */
+#    define INCBIN_TYPE(NAME)    ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", %object\n"
+#  elif defined(__MINGW32__) || defined(__MINGW64__)
+/* Mingw doesn't support this directive either */
+#    define INCBIN_TYPE(NAME)
+#  else
+/* It's safe to use `@' on other architectures */
+#    define INCBIN_TYPE(NAME)    ".type " INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME ", @object\n"
+#  endif
+#  define INCBIN_BYTE            ".byte "
+#endif
+
+/* List of style types used for symbol names */
+#define INCBIN_STYLE_CAMEL 0
+#define INCBIN_STYLE_SNAKE 1
+
+/**
+ * @brief Specify the prefix to use for symbol names.
+ *
+ * @note By default this is "g".
+ *
+ * @code
+ * #define INCBIN_PREFIX incbin
+ * #include "incbin.h"
+ * INCBIN(Foo, "foo.txt");
+ *
+ * // Now you have the following symbols instead:
+ * // const unsigned char incbinFoo<data>[];
+ * // const unsigned char *const incbinFoo<end>;
+ * // const unsigned int incbinFoo<size>;
+ * @endcode
+ */
+#if !defined(INCBIN_PREFIX)
+#  define INCBIN_PREFIX g
+#endif
+
+/**
+ * @brief Specify the style used for symbol names.
+ *
+ * Possible options are
+ * - INCBIN_STYLE_CAMEL "CamelCase"
+ * - INCBIN_STYLE_SNAKE "snake_case"
+ *
+ * @note By default this is INCBIN_STYLE_CAMEL
+ *
+ * @code
+ * #define INCBIN_STYLE INCBIN_STYLE_SNAKE
+ * #include "incbin.h"
+ * INCBIN(foo, "foo.txt");
+ *
+ * // Now you have the following symbols:
+ * // const unsigned char <prefix>foo_data[];
+ * // const unsigned char *const <prefix>foo_end;
+ * // const unsigned int <prefix>foo_size;
+ * @endcode
+ */
+#if !defined(INCBIN_STYLE)
+#  define INCBIN_STYLE INCBIN_STYLE_CAMEL
+#endif
+
+/* Style lookup tables */
+#define INCBIN_STYLE_0_DATA Data
+#define INCBIN_STYLE_0_END End
+#define INCBIN_STYLE_0_SIZE Size
+#define INCBIN_STYLE_1_DATA _data
+#define INCBIN_STYLE_1_END _end
+#define INCBIN_STYLE_1_SIZE _size
+
+/* Style lookup: returning identifier */
+#define INCBIN_STYLE_IDENT(TYPE) \
+    INCBIN_CONCATENATE( \
+        INCBIN_STYLE_, \
+        INCBIN_CONCATENATE( \
+            INCBIN_EVAL(INCBIN_STYLE), \
+            INCBIN_CONCATENATE(_, TYPE)))
+
+/* Style lookup: returning string literal */
+#define INCBIN_STYLE_STRING(TYPE) \
+    INCBIN_STRINGIZE( \
+        INCBIN_STYLE_IDENT(TYPE)) \
+
+/* Generate the global labels by indirectly invoking the macro with our style
+ * type and concatenating the name against them. */
+#define INCBIN_GLOBAL_LABELS(NAME, TYPE) \
+    INCBIN_INVOKE( \
+        INCBIN_GLOBAL, \
+        INCBIN_CONCATENATE( \
+            NAME, \
+            INCBIN_INVOKE( \
+                INCBIN_STYLE_IDENT, \
+                TYPE))) \
+    INCBIN_INVOKE( \
+        INCBIN_TYPE, \
+        INCBIN_CONCATENATE( \
+            NAME, \
+            INCBIN_INVOKE( \
+                INCBIN_STYLE_IDENT, \
+                TYPE)))
+
+/**
+ * @brief Externally reference binary data included in another translation unit.
+ *
+ * Produces three external symbols that reference the binary data included in
+ * another translation unit.
+ *
+ * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
+ * "Data", as well as "End" and "Size" after. An example is provided below.
+ *
+ * @param TYPE Optional array type. Omitting this picks a default of `unsigned char`.
+ * @param NAME The name given for the binary data
+ *
+ * @code
+ * INCBIN_EXTERN(Foo);
+ *
+ * // Now you have the following symbols:
+ * // extern const unsigned char <prefix>Foo<data>[];
+ * // extern const unsigned char *const <prefix>Foo<end>;
+ * // extern const unsigned int <prefix>Foo<size>;
+ * @endcode
+ * 
+ * You may specify a custom optional data type as well as the first argument.
+ * @code
+ * INCBIN_EXTERN(custom_type, Foo);
+ * 
+ * // Now you have the following symbols:
+ * // extern const custom_type <prefix>Foo<data>[];
+ * // extern const custom_type *const <prefix>Foo<end>;
+ * // extern const unsigned int <prefix>Foo<size>;
+ * @endcode
+ */
+#define INCBIN_EXTERN(...) \
+    INCBIN_CONCATENATE(INCBIN_EXTERN_, INCBIN_VA_ARGC(__VA_ARGS__))(__VA_ARGS__)
+#define INCBIN_EXTERN_1(NAME, ...) \
+    INCBIN_EXTERN_2(unsigned char, NAME)
+#define INCBIN_EXTERN_2(TYPE, NAME) \
+    INCBIN_EXTERNAL const INCBIN_ALIGN TYPE \
+        INCBIN_CONCATENATE( \
+            INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
+            INCBIN_STYLE_IDENT(DATA))[]; \
+    INCBIN_EXTERNAL const INCBIN_ALIGN TYPE *const \
+    INCBIN_CONCATENATE( \
+        INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
+        INCBIN_STYLE_IDENT(END)); \
+    INCBIN_EXTERNAL const unsigned int \
+        INCBIN_CONCATENATE( \
+            INCBIN_CONCATENATE(INCBIN_PREFIX, NAME), \
+            INCBIN_STYLE_IDENT(SIZE))
+
+/**
+ * @brief Externally reference textual data included in another translation unit.
+ *
+ * Produces three external symbols that reference the textual data included in
+ * another translation unit.
+ *
+ * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
+ * "Data", as well as "End" and "Size" after. An example is provided below.
+ *
+ * @param NAME The name given for the textual data
+ *
+ * @code
+ * INCBIN_EXTERN(Foo);
+ *
+ * // Now you have the following symbols:
+ * // extern const char <prefix>Foo<data>[];
+ * // extern const char *const <prefix>Foo<end>;
+ * // extern const unsigned int <prefix>Foo<size>;
+ * @endcode
+ */
+#define INCTXT_EXTERN(NAME) \
+    INCBIN_EXTERN_2(char, NAME)
+
+/**
+ * @brief Include a binary file into the current translation unit.
+ *
+ * Includes a binary file into the current translation unit, producing three symbols
+ * for objects that encode the data and size respectively.
+ *
+ * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
+ * "Data", as well as "End" and "Size" after. An example is provided below.
+ *
+ * @param TYPE Optional array type. Omitting this picks a default of `unsigned char`.
+ * @param NAME The name to associate with this binary data (as an identifier.)
+ * @param FILENAME The file to include (as a string literal.)
+ *
+ * @code
+ * INCBIN(Icon, "icon.png");
+ *
+ * // Now you have the following symbols:
+ * // const unsigned char <prefix>Icon<data>[];
+ * // const unsigned char *const <prefix>Icon<end>;
+ * // const unsigned int <prefix>Icon<size>;
+ * @endcode
+ * 
+ * You may specify a custom optional data type as well as the first argument.
+ * These macros are specialized by arity.
+ * @code
+ * INCBIN(custom_type, Icon, "icon.png");
+ *
+ * // Now you have the following symbols:
+ * // const custom_type <prefix>Icon<data>[];
+ * // const custom_type *const <prefix>Icon<end>;
+ * // const unsigned int <prefix>Icon<size>;
+ * @endcode
+ *
+ * @warning This must be used in global scope
+ * @warning The identifiers may be different if INCBIN_STYLE is not default
+ *
+ * To externally reference the data included by this in another translation unit
+ * please @see INCBIN_EXTERN.
+ */
+#ifdef _MSC_VER
+#  define INCBIN(NAME, FILENAME) \
+      INCBIN_EXTERN(NAME)
+#else
+#  define INCBIN(...) \
+     INCBIN_CONCATENATE(INCBIN_, INCBIN_VA_ARGC(__VA_ARGS__))(__VA_ARGS__)
+#  if defined(__GNUC__)
+#    define INCBIN_1(...) _Pragma("GCC error \"Single argument INCBIN not allowed\"")
+#  elif defined(__clang__)
+#    define INCBIN_1(...) _Pragma("clang error \"Single argument INCBIN not allowed\"")
+#  else
+#    define INCBIN_1(...) /* Cannot do anything here */
+#  endif
+#  define INCBIN_2(NAME, FILENAME) \
+      INCBIN_3(unsigned char, NAME, FILENAME)
+#  define INCBIN_3(TYPE, NAME, FILENAME) INCBIN_COMMON(TYPE, NAME, FILENAME, /* No terminator for binary data */)
+#  define INCBIN_COMMON(TYPE, NAME, FILENAME, TERMINATOR) \
+    __asm__(INCBIN_SECTION \
+            INCBIN_GLOBAL_LABELS(NAME, DATA) \
+            INCBIN_ALIGN_HOST \
+            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) ":\n" \
+            INCBIN_MACRO " \"" FILENAME "\"\n" \
+                TERMINATOR \
+            INCBIN_GLOBAL_LABELS(NAME, END) \
+            INCBIN_ALIGN_BYTE \
+            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) ":\n" \
+                INCBIN_BYTE "1\n" \
+            INCBIN_GLOBAL_LABELS(NAME, SIZE) \
+            INCBIN_ALIGN_HOST \
+            INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(SIZE) ":\n" \
+                INCBIN_INT INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(END) " - " \
+                           INCBIN_MANGLE INCBIN_STRINGIZE(INCBIN_PREFIX) #NAME INCBIN_STYLE_STRING(DATA) "\n" \
+            INCBIN_ALIGN_HOST \
+            ".text\n" \
+    ); \
+    INCBIN_EXTERN(TYPE, NAME)
+#endif
+
+/**
+ * @brief Include a textual file into the current translation unit.
+ * 
+ * This behaves the same as INCBIN except it produces char compatible arrays
+ * and implicitly adds a null-terminator byte, thus the size of data included
+ * by this is one byte larger than that of INCBIN.
+ *
+ * Includes a textual file into the current translation unit, producing three
+ * symbols for objects that encode the data and size respectively.
+ *
+ * The symbol names are a concatenation of `INCBIN_PREFIX' before *NAME*; with
+ * "Data", as well as "End" and "Size" after. An example is provided below.
+ *
+ * @param NAME The name to associate with this binary data (as an identifier.)
+ * @param FILENAME The file to include (as a string literal.)
+ *
+ * @code
+ * INCTXT(Readme, "readme.txt");
+ *
+ * // Now you have the following symbols:
+ * // const char <prefix>Readme<data>[];
+ * // const char *const <prefix>Readme<end>;
+ * // const unsigned int <prefix>Readme<size>;
+ * @endcode
+ *
+ * @warning This must be used in global scope
+ * @warning The identifiers may be different if INCBIN_STYLE is not default
+ *
+ * To externally reference the data included by this in another translation unit
+ * please @see INCBIN_EXTERN.
+ */
+#if defined(_MSC_VER)
+#  define INCTXT(NAME, FILENAME) \
+     INCBIN_EXTERN(NAME)
+#else
+#  define INCTXT(NAME, FILENAME) \
+     INCBIN_COMMON(char, NAME, FILENAME, INCBIN_BYTE "0\n")
+#endif
+
+#endif

SmartCrop/intelligentroi.cpp

@@ -0,0 +1,128 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include "intelligentroi.h"
+
+#include <opencv2/imgproc.hpp>
+
+#include "utils.h"
+#include "log.h"
+
+bool InteligentRoi::compPointPrio(const std::pair<cv::Point2i, int>& a, const std::pair<cv::Point2i, int>& b, const cv::Point2i& center)
+{
+	if(a.second != b.second)
+		return a.second > b.second;
+
+	double distA = pointDist(a.first, center);
+	double distB = pointDist(b.first, center);
+
+	return distA < distB;
+}
+
+void InteligentRoi::slideRectToPoint(cv::Rect& rect, const cv::Point2i& point)
+{
+	if(!pointInRect(point, rect))
+	{
+		if(point.x < rect.x)
+			rect.x = point.x;
+		else if(point.x > rect.x+rect.width)
+			rect.x = point.x-rect.width;
+		if(point.y < rect.y)
+			rect.y = point.y;
+		else if(point.y > rect.y+rect.height)
+			rect.y = point.y-rect.height;
+	}
+}
+
+cv::Rect InteligentRoi::maxRect(bool& incompleate, const cv::Size2i& imageSize, std::vector<std::pair<cv::Point2i, int>> mustInclude)
+{
+	incompleate = false;
+	int diameter = std::min(imageSize.height, imageSize.width);
+	cv::Point2i point(imageSize.width/2, imageSize.height/2);
+	cv::Rect candiate(point.x-diameter/2, point.y-diameter/2, diameter, diameter);
+
+	std::sort(mustInclude.begin(), mustInclude.end(),
+		[&point](const std::pair<cv::Point2i, int>& a, const std::pair<cv::Point2i, int>& b){return compPointPrio(a, b, point);});
+
+	while(true)
+	{
+		cv::Rect includeRect = rectFromPoints(mustInclude);
+		if(includeRect.width-2 > diameter || includeRect.height-2 > diameter)
+		{
+			incompleate = true;
+			slideRectToPoint(candiate, mustInclude.back().first);
+			mustInclude.pop_back();
+			Log(Log::DEBUG)<<"cant fill";
+			for(const std::pair<cv::Point2i, int>& mipoint : mustInclude)
+				Log(Log::DEBUG)<<mipoint.first<<' '<<pointDist(mipoint.first, point)<<' '<<mipoint.second;
+		}
+		else
+		{
+			break;
+		}
+	}
+
+	for(const std::pair<cv::Point2i, int>& includePoint : mustInclude)
+		slideRectToPoint(candiate, includePoint.first);
+
+	if(candiate.x < 0)
+		candiate.x = 0;
+	if(candiate.y < 0)
+		candiate.y = 0;
+	if(candiate.x+candiate.width > imageSize.width)
+		candiate.width = imageSize.width-candiate.x;
+	if(candiate.y+candiate.height > imageSize.height)
+		candiate.height = imageSize.height-candiate.y;
+
+	return candiate;
+}
+
+InteligentRoi::InteligentRoi(const Yolo& yolo)
+{
+	personId = yolo.getClassForStr("person");
+}
+
+bool InteligentRoi::getCropRectangle(cv::Rect& out, const std::vector<Yolo::Detection>& detections, const cv::Size2i& imageSize)
+{
+	std::vector<std::pair<cv::Point2i, int>> corners;
+	for(size_t i = 0; i < detections.size(); ++i)
+	{
+		int priority = detections[i].priority;
+		if(detections[i].class_id == personId)
+		{
+			corners.push_back({detections[i].box.tl()+cv::Point2i(detections[i].box.width/2, 0), priority+2});
+			corners.push_back({detections[i].box.tl(), priority+1});
+			corners.push_back({detections[i].box.br(), priority});
+			corners.push_back({detections[i].box.tl()+cv::Point2i(detections[i].box.width, 0), priority+1});
+			corners.push_back({detections[i].box.br()+cv::Point2i(0-detections[i].box.width, 0), priority});
+		}
+		else
+		{
+			corners.push_back({detections[i].box.tl(), priority});
+			corners.push_back({detections[i].box.br(), priority});
+			corners.push_back({detections[i].box.tl()+cv::Point2i(detections[i].box.width, 0), priority});
+			corners.push_back({detections[i].box.br()+cv::Point2i(0-detections[i].box.width, 0), priority});
+		}
+	}
+
+	bool incompleate;
+	out = maxRect(incompleate, imageSize, corners);
+	return incompleate;
+}

SmartCrop/intelligentroi.h

@@ -0,0 +1,37 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include <opencv2/imgproc.hpp>
+
+#include "yolo.h"
+
+class InteligentRoi
+{
+private:
+	int personId;
+	static bool compPointPrio(const std::pair<cv::Point2i, int>& a, const std::pair<cv::Point2i, int>& b, const cv::Point2i& center);
+	static void slideRectToPoint(cv::Rect& rect, const cv::Point2i& point);
+	static cv::Rect maxRect(bool& incompleate, const cv::Size2i& imageSize, std::vector<std::pair<cv::Point2i, int>> mustInclude = {});
+
+public:
+	InteligentRoi(const Yolo& yolo);
+	bool getCropRectangle(cv::Rect& out, const std::vector<Yolo::Detection>& detections, const cv::Size2i& imageSize);
+};

SmartCrop/log.cpp

@@ -0,0 +1,63 @@
+/**
+* Lubricant Detecter
+* Copyright (C) 2021 Carl Klemm
+*
+* This program is free software; you can redistribute it and/or
+* modify it under the terms of the GNU General Public License
+* version 3 as published by the Free Software Foundation.
+*
+* This program is distributed in the hope that it will be useful,
+* but WITHOUT ANY WARRANTY; without even the implied warranty of
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+* GNU General Public License for more details.
+*
+* You should have received a copy of the GNU General Public License
+* along with this program; if not, write to the
+* Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+* Boston, MA  02110-1301, USA.
+*/
+
+#include "log.h"
+
+Log::Log(Level type, bool endlineI): endline(endlineI)
+{
+	msglevel = type;
+	if(headers)
+	{
+		operator << ("["+getLabel(type)+"] ");
+	}
+}
+
+Log::~Log()
+{
+	if(opened && endline)
+	{
+		std::cout<<'\n';
+	}
+	opened = false;
+}
+
+
+std::string Log::getLabel(Level level)
+{
+	std::string label;
+	switch(level)
+	{
+		case DEBUG:
+			label = "DEBUG";
+			break;
+		case INFO:
+			label = "INFO ";
+			break;
+		case WARN:
+			label = "WARN ";
+			break;
+		case ERROR:
+			label = "ERROR";
+			break;
+	}
+	return label;
+}
+
+bool Log::headers = false;
+Log::Level Log::level = WARN;

SmartCrop/log.h

@@ -0,0 +1,64 @@
+/**
+* eisgenerator
+* Copyright (C) 2021 Carl Klemm
+*
+* This program is free software; you can redistribute it and/or
+* modify it under the terms of the GNU General Public License
+* version 3 as published by the Free Software Foundation.
+*
+* This program is distributed in the hope that it will be useful,
+* but WITHOUT ANY WARRANTY; without even the implied warranty of
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+* GNU General Public License for more details.
+*
+* You should have received a copy of the GNU General Public License
+* along with this program; if not, write to the
+* Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+* Boston, MA  02110-1301, USA.
+*/
+
+#pragma once
+#include <iostream>
+#include <string>
+
+class Log
+{
+public:
+
+	enum Level
+	{
+		DEBUG,
+		INFO,
+		WARN,
+		ERROR
+	};
+
+private:
+	bool opened = false;
+	Level msglevel = DEBUG;
+	bool endline = true;
+
+	std::string getLabel(Level level);
+
+public:
+
+	static bool headers;
+	static Level level;
+
+	Log() {}
+	Log(Level type, bool endlineI = true);
+	~Log();
+
+	template<class T> Log &operator<<(const T &msg)
+	{
+		if(msglevel >= level)
+		{
+			if(msglevel == ERROR)
+				std::cerr<<msg;
+			else
+				std::cout<<msg;
+			opened = true;
+		}
+		return *this;
+	}
+};

SmartCrop/main.cpp

@@ -0,0 +1,460 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include <filesystem>
+#include <iostream>
+#include <opencv2/core.hpp>
+#include <opencv2/core/types.hpp>
+#include <opencv2/imgproc.hpp>
+#include <opencv2/highgui.hpp>
+#include <algorithm>
+#include <execution>
+#include <string>
+#include <vector>
+#include <numeric>
+
+#include "yolo.h"
+#include "log.h"
+#include "options.h"
+#include "utils.h"
+#include "intelligentroi.h"
+#include "seamcarving.h"
+#include "facerecognizer.h"
+
+const Yolo::Detection* pointInDetectionHoriz(int x, const std::vector<Yolo::Detection>& detections, const Yolo::Detection* ignore = nullptr)
+{
+	const Yolo::Detection* inDetection = nullptr;
+	for(const Yolo::Detection& detection : detections)
+	{
+		if(ignore && ignore == &detection)
+			continue;
+
+		if(detection.box.x <= x && detection.box.x+detection.box.width >= x)
+		{
+			if(!inDetection || detection.box.br().x > inDetection->box.br().x)
+			inDetection = &detection;
+		}
+	}
+	return inDetection;
+}
+
+bool findRegionEndpointHoriz(int& x, const std::vector<Yolo::Detection>& detections, int imgSizeX)
+{
+	const Yolo::Detection* inDetection = pointInDetectionHoriz(x, detections);
+
+	Log(Log::DEBUG, false)<<__func__<<" point "<<x;
+
+	if(!inDetection)
+	{
+		const Yolo::Detection* closest = nullptr;
+		for(const Yolo::Detection& detection : detections)
+		{
+			if(detection.box.x > x)
+			{
+				if(closest == nullptr || detection.box.x-x > closest->box.x-x)
+					closest = &detection;
+			}
+		}
+		if(closest)
+			x = closest->box.x;
+		else
+			x = imgSizeX;
+
+		Log(Log::DEBUG)<<" is not in any box and will be moved to "<<x<<" where the closest box ("<<(closest ? closest->className : "null")<<") is";
+		return false;
+	}
+	else
+	{
+		x = inDetection->box.br().x;
+		Log(Log::DEBUG, false)<<" is in a box and will be moved to its end "<<x<<" where ";
+		const Yolo::Detection* candidateDetection = pointInDetectionHoriz(x, detections, inDetection);
+		if(candidateDetection && candidateDetection->box.br().x > x)
+		{
+			Log(Log::DEBUG)<<"it is again in a box";
+			return findRegionEndpointHoriz(x, detections, imgSizeX);
+		}
+		else
+		{
+			Log(Log::DEBUG)<<"it is not in a box";
+			return true;
+		}
+	}
+}
+
+std::vector<std::pair<cv::Mat, bool>> cutImageIntoHorzRegions(cv::Mat& image, const std::vector<Yolo::Detection>& detections)
+{
+	std::vector<std::pair<cv::Mat, bool>> out;
+
+	std::cout<<__func__<<' '<<image.cols<<'x'<<image.rows<<std::endl;
+
+	for(int x = 0; x < image.cols; ++x)
+	{
+		int start = x;
+		bool frozen = findRegionEndpointHoriz(x, detections, image.cols);
+
+		int width = x-start;
+		if(x < image.cols)
+			++width;
+		cv::Rect rect(start, 0, width, image.rows);
+		Log(Log::DEBUG)<<__func__<<" region\t"<<rect;
+		cv::Mat slice = image(rect);
+		out.push_back({slice, frozen});
+	}
+
+	return out;
+}
+
+cv::Mat assembleFromSlicesHoriz(const std::vector<std::pair<cv::Mat, bool>>& slices)
+{
+	assert(!slices.empty());
+
+	int cols = 0;
+	for(const std::pair<cv::Mat, bool>& slice : slices)
+		cols += slice.first.cols;
+
+
+	cv::Mat image(cols, slices[0].first.rows, slices[0].first.type());
+	Log(Log::DEBUG)<<__func__<<' '<<image.size()<<' '<<cols<<' '<<slices[0].first.rows;
+
+	int col = 0;
+	for(const std::pair<cv::Mat, bool>& slice : slices)
+	{
+		cv::Rect rect(col, 0, slice.first.cols, slice.first.rows);
+		Log(Log::DEBUG)<<__func__<<' '<<rect;
+		slice.first.copyTo(image(rect));
+		col += slice.first.cols-1;
+	}
+
+	return image;
+}
+
+void transposeRect(cv::Rect& rect)
+{
+	int x = rect.x;
+	rect.x = rect.y;
+	rect.y = x;
+
+	int width = rect.width;
+	rect.width = rect.height;
+	rect.height = width;
+}
+
+bool seamCarveResize(cv::Mat& image, std::vector<Yolo::Detection> detections, double targetAspectRatio = 1.0)
+{
+	detections.erase(std::remove_if(detections.begin(), detections.end(), [](const Yolo::Detection& detection){return detection.priority < 3;}), detections.end());
+
+	double aspectRatio = image.cols/static_cast<double>(image.rows);
+
+	Log(Log::DEBUG)<<"Image size "<<image.size()<<" aspect ratio "<<aspectRatio<<" target aspect ratio "<<targetAspectRatio;
+
+	bool vertical = false;
+	if(aspectRatio > targetAspectRatio)
+		vertical = true;
+
+	int requiredLines = 0;
+	if(!vertical)
+		requiredLines = image.rows*targetAspectRatio - image.cols;
+	else
+		requiredLines = image.cols/targetAspectRatio - image.rows;
+
+	Log(Log::DEBUG)<<__func__<<' '<<requiredLines<<" lines are required in "<<(vertical ? "vertical" : "horizontal")<<" direction";
+
+	if(vertical)
+	{
+		cv::transpose(image, image);
+		for(Yolo::Detection& detection : detections)
+			transposeRect(detection.box);
+	}
+
+	std::vector<std::pair<cv::Mat, bool>> slices = cutImageIntoHorzRegions(image, detections);
+	Log(Log::DEBUG)<<"Image has "<<slices.size()<<" slices:";
+	int totalResizableSize = 0;
+	for(const std::pair<cv::Mat, bool>& slice : slices)
+	{
+		Log(Log::DEBUG)<<"a "<<(slice.second ? "frozen" : "unfrozen")<<" slice of size "<<slice.first.cols;
+		if(!slice.second)
+			totalResizableSize += slice.first.cols;
+	}
+
+	if(totalResizableSize < requiredLines+1)
+	{
+		Log(Log::WARN)<<"Unable to seam carve as there are only "<<totalResizableSize<<" unfrozen cols";
+		if(vertical)
+			cv::transpose(image, image);
+		return false;
+	}
+
+	std::vector<int> seamsForSlice(slices.size(), 0);
+	for(size_t i = 0; i < slices.size(); ++i)
+	{
+		if(!slices[i].second)
+			seamsForSlice[i] = (static_cast<double>(slices[i].first.cols)/totalResizableSize)*requiredLines;
+	}
+
+	int residual = requiredLines - std::accumulate(seamsForSlice.begin(), seamsForSlice.end(), decltype(seamsForSlice)::value_type(0));;
+	for(ssize_t i = slices.size()-1; i >= 0; --i)
+	{
+		if(!slices[i].second)
+		{
+			seamsForSlice[i] += residual;
+			break;
+		}
+	}
+
+	for(size_t i = 0; i < slices.size(); ++i)
+	{
+		if(seamsForSlice[i] != 0)
+		{
+			bool ret = SeamCarving::strechImage(slices[i].first, seamsForSlice[i], true);
+			if(!ret)
+			{
+				if(vertical)
+					transpose(image, image);
+				return false;
+			}
+		}
+	}
+
+	image = assembleFromSlicesHoriz(slices);
+
+	if(vertical)
+		cv::transpose(image, image);
+
+	return true;
+}
+
+void drawDebugInfo(cv::Mat &image, const cv::Rect& rect, const std::vector<Yolo::Detection>& detections)
+{
+	for(const Yolo::Detection& detection : detections)
+	{
+		cv::rectangle(image, detection.box, detection.color, 3);
+		std::string label = detection.className + ' ' + std::to_string(detection.confidence).substr(0, 4) + ' ' + std::to_string(detection.priority);
+		cv::Size labelSize = cv::getTextSize(label, cv::FONT_HERSHEY_DUPLEX, 1, 1, 0);
+		cv::Rect textBox(detection.box.x, detection.box.y - 40, labelSize.width + 10, labelSize.height + 20);
+		cv::rectangle(image, textBox, detection.color, cv::FILLED);
+		cv::putText(image, label, cv::Point(detection.box.x + 5, detection.box.y - 10), cv::FONT_HERSHEY_DUPLEX, 1, cv::Scalar(0, 0, 0), 1, 0);
+	}
+
+	cv::rectangle(image, rect, cv::Scalar(0, 0, 255), 8);
+}
+
+static void reduceSize(cv::Mat& image, const cv::Size& targetSize)
+{
+	int longTargetSize = std::max(targetSize.width, targetSize.height)*2;
+	if(std::max(image.cols, image.rows) > longTargetSize)
+	{
+		if(image.cols > image.rows)
+		{
+			double ratio = static_cast<double>(longTargetSize)/image.cols;
+			cv::resize(image, image, {longTargetSize, static_cast<int>(image.rows*ratio)}, 0, 0, ratio < 1 ? cv::INTER_AREA : cv::INTER_CUBIC);
+		}
+		else
+		{
+			double ratio = static_cast<double>(longTargetSize)/image.rows;
+			cv::resize(image, image, {static_cast<int>(image.cols*ratio), longTargetSize}, 0, 0, ratio < 1 ? cv::INTER_AREA : cv::INTER_CUBIC);
+		}
+	}
+}
+
+void pipeline(const std::filesystem::path& path, const Config& config, Yolo& yolo, FaceRecognizer* recognizer,
+	std::mutex& reconizerMutex, const std::filesystem::path& debugOutputPath)
+{
+	InteligentRoi intRoi(yolo);
+	cv::Mat image = cv::imread(path);
+	if(!image.data)
+	{
+		Log(Log::WARN)<<"could not load image "<<path<<" skipping";
+		return;
+	}
+
+	reduceSize(image, config.targetSize);
+
+	std::vector<Yolo::Detection> detections = yolo.runInference(image);
+
+	Log(Log::DEBUG)<<"Got "<<detections.size()<<" detections for "<<path;
+	for(Yolo::Detection& detection : detections)
+	{
+		bool hasmatch = false;
+		if(recognizer && detection.className == "person")
+		{
+			cv::Mat person = image(detection.box);
+			reconizerMutex.lock();
+			FaceRecognizer::Detection match = recognizer->isMatch(person);
+			reconizerMutex.unlock();
+			if(match.person >= 0)
+			{
+				detection.priority += 10;
+				hasmatch = true;
+				detections.push_back({0, "Face", match.confidence, 20, {255, 0, 0}, match.rect});
+			}
+		}
+		Log(Log::DEBUG)<<detection.class_id<<": "<<detection.className<<" at "<<detection.box<<" with prio "<<detection.priority<<(hasmatch ? " has match" : "");
+	}
+
+	cv::Rect crop;
+	bool incompleate = intRoi.getCropRectangle(crop, detections, image.size());
+
+	if(config.seamCarving && incompleate)
+	{
+		bool ret = seamCarveResize(image, detections, config.targetSize.aspectRatio());
+		if(ret && image.size().aspectRatio() != config.targetSize.aspectRatio())
+		{
+			detections = yolo.runInference(image);
+		}
+	}
+
+	cv::Mat croppedImage;
+
+	if(image.size().aspectRatio() != config.targetSize.aspectRatio() && incompleate)
+	{
+		intRoi.getCropRectangle(crop, detections, image.size());
+
+		if(config.debug)
+		{
+			cv::Mat debugImage = image.clone();
+			drawDebugInfo(debugImage, crop, detections);
+			bool ret = cv::imwrite(debugOutputPath/path.filename(), debugImage);
+			if(!ret)
+				Log(Log::WARN)<<"could not save debug image to "<<debugOutputPath/path.filename()<<" skipping";
+		}
+
+		croppedImage = image(crop);
+	}
+	else if(!incompleate)
+	{
+		croppedImage = image(crop);
+	}
+	else
+	{
+		croppedImage = image;
+	}
+
+	cv::Mat resizedImage;
+	cv::resize(croppedImage, resizedImage, config.targetSize, 0, 0, cv::INTER_CUBIC);
+	bool ret = cv::imwrite(config.outputDir/path.filename(), resizedImage);
+	if(!ret)
+		Log(Log::WARN)<<"could not save image to "<<config.outputDir/path.filename()<<" skipping";
+}
+
+void threadFn(const std::vector<std::filesystem::path>& images, const Config& config, FaceRecognizer* recognizer,
+		std::mutex& reconizerMutex, const std::filesystem::path& debugOutputPath)
+{
+	Yolo yolo(config.modelPath, {640, 480}, config.classesPath, false);
+	for(std::filesystem::path path : images)
+		pipeline(path, config, yolo, recognizer, reconizerMutex, debugOutputPath);
+}
+
+template<typename T>
+std::vector<std::vector<T>> splitVector(const std::vector<T>& vec, size_t parts)
+{
+	std::vector<std::vector<T>> out;
+
+	size_t length = vec.size()/parts;
+	size_t remain = vec.size() % parts;
+
+	size_t begin = 0;
+	size_t end = 0;
+
+	for (size_t i = 0; i < std::min(parts, vec.size()); ++i)
+	{
+		end += (remain > 0) ? (length + !!(remain--)) : length;
+		out.push_back(std::vector<T>(vec.begin() + begin, vec.begin() + end));
+		begin = end;
+	}
+
+	return out;
+}
+
+int main(int argc, char* argv[])
+{
+	Log::level = Log::INFO;
+
+	Config config;
+	argp_parse(&argp, argc, argv, 0, 0, &config);
+
+	if(config.outputDir.empty())
+	{
+		Log(Log::ERROR)<<"a output path \"-o\" is required";
+		return 1;
+	}
+
+	if(config.imagePaths.empty())
+	{
+		Log(Log::ERROR)<<"at least one input image or directory is required";
+		return 1;
+	}
+
+	std::vector<std::filesystem::path> imagePaths;
+
+	for(const std::filesystem::path& path : config.imagePaths)
+		getImageFiles(path, imagePaths);
+
+	Log(Log::DEBUG)<<"Images:";
+	for(const::std::filesystem::path& path: imagePaths)
+		Log(Log::DEBUG)<<path;
+
+	if(imagePaths.empty())
+	{
+		Log(Log::ERROR)<<"no image was found\n";
+		return 1;
+	}
+
+	if(!std::filesystem::exists(config.outputDir))
+	{
+		if(!std::filesystem::create_directory(config.outputDir))
+		{
+			Log(Log::ERROR)<<"could not create directory at "<<config.outputDir;
+			return 1;
+		}
+	}
+
+	std::filesystem::path debugOutputPath(config.outputDir/"debug");
+	if(config.debug)
+	{
+		if(!std::filesystem::exists(debugOutputPath))
+			std::filesystem::create_directory(debugOutputPath);
+	}
+
+	FaceRecognizer* recognizer = nullptr;
+	std::mutex recognizerMutex;
+	if(!config.focusPersonImage.empty())
+	{
+		cv::Mat personImage = cv::imread(config.focusPersonImage);
+		if(personImage.empty())
+		{
+			Log(Log::ERROR)<<"Could not load image from "<<config.focusPersonImage;
+			return 1;
+		}
+		recognizer = new FaceRecognizer();
+		recognizer->addReferances({personImage});
+		recognizer->setThreshold(config.threshold);
+	}
+
+	std::vector<std::thread> threads;
+	std::vector<std::vector<std::filesystem::path>> imagePathParts = splitVector(imagePaths, std::thread::hardware_concurrency());
+
+	for(size_t i = 0; i < imagePathParts.size(); ++i)
+		threads.push_back(std::thread(threadFn, imagePathParts[i], std::ref(config),  recognizer, std::ref(recognizerMutex), std::ref(debugOutputPath)));
+
+	for(std::thread& thread : threads)
+		thread.join();
+
+	return 0;
+}

SmartCrop/options.h

@@ -0,0 +1,117 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+#include <string>
+#include <vector>
+#include <argp.h>
+#include <iostream>
+#include <filesystem>
+#include <opencv2/core/types.hpp>
+#include "log.h"
+
+const char *argp_program_version = "AIImagePreprocesses";
+const char *argp_program_bug_address = "<carl@uvos.xyz>";
+static char doc[] = "Application that trainsforms images into formats, sizes and aspect ratios required for ai training";
+static char args_doc[] = "FILE(S)";
+
+static struct argp_option options[] =
+{
+  {"verbose",		'v', 0,				0,	"Show debug messages" },
+  {"quiet", 		'q', 0,				0,	"only output data" },
+  {"model", 		'm', "[FILENAME]",	0,	"YoloV8 model to use for detection" },
+  {"classes", 		'c', "[FILENAME]",	0,	"classes text file to use" },
+  {"out",	 		'o', "[DIRECTORY]",	0,	"directory whre images are to be saved" },
+  {"debug", 		'd', 0,				0,	"output debug images" },
+  {"seam-carving", 	's', 0,				0,	"use seam carving to change image aspect ratio instead of croping"},
+  {"size", 			'z', "[PIXELS]",	0,	"target output size, default: 512"},
+  {"focus-person",	'f', "[FILENAME]",	0,	"a file name to an image of a person that the crop should focus on"},
+  {"person-threshold",	't', "[NUMBER]",	0,	"the threshold at witch to consider a person matched, defaults to 0.363"},
+  {0}
+};
+
+struct Config
+{
+	std::vector<std::filesystem::path> imagePaths;
+	std::filesystem::path modelPath;
+	std::filesystem::path classesPath;
+	std::filesystem::path outputDir;
+	std::filesystem::path focusPersonImage;
+	bool seamCarving = false;
+	bool debug = false;
+	double threshold = 0.363;
+	cv::Size targetSize = cv::Size(512, 512);
+};
+
+static error_t parse_opt (int key, char *arg, struct argp_state *state)
+{
+	Config *config = reinterpret_cast<Config*>(state->input);
+	try
+	{
+		switch (key)
+		{
+		case 'q':
+			Log::level = Log::ERROR;
+			break;
+		case 'v':
+			Log::level = Log::DEBUG;
+			break;
+		case 'm':
+			config->modelPath = arg;
+			break;
+		case 'c':
+			config->classesPath = arg;
+			break;
+		case 'd':
+			config->debug = true;
+			break;
+		case 'o':
+			config->outputDir.assign(arg);
+			break;
+		case 's':
+			config->seamCarving = true;
+			break;
+		case 'f':
+			config->focusPersonImage = arg;
+			break;
+		case 't':
+			config->threshold = std::atof(arg);
+			break;
+		case 'z':
+		{
+			int x = std::stoi(arg);
+			config->targetSize = cv::Size(x, x);
+			break;
+		}
+		case ARGP_KEY_ARG:
+			config->imagePaths.push_back(arg);
+			break;
+		default:
+			return ARGP_ERR_UNKNOWN;
+		}
+	}
+	catch(const std::invalid_argument& ex)
+	{
+		std::cout<<arg<<" passed for argument -"<<static_cast<char>(key)<<" is not a valid number.\n";
+		return ARGP_KEY_ERROR;
+	}
+	return 0;
+}
+
+static struct argp argp = {options, parse_opt, args_doc, doc};

SmartCrop/readfile.h

@@ -0,0 +1,35 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+#include <string>
+#include <filesystem>
+#include <fstream>
+#include <stdexcept>
+#include <sstream>
+
+inline std::string readFile(const std::filesystem::path& path)
+{
+	std::ifstream file(path);
+	if(!file.is_open())
+		throw std::runtime_error(std::string("could not open file ") + path.string());
+	std::stringstream ss;
+	ss<<file.rdbuf();
+	return ss.str();
+}

SmartCrop/seamcarving.cpp

@@ -0,0 +1,376 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include "seamcarving.h"
+
+#include <opencv2/imgcodecs.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc.hpp>
+#include <iostream>
+#include <filesystem>
+#include <cfloat>
+#include <vector>
+#include "log.h"
+
+bool SeamCarving::strechImage(cv::Mat& image, int seams, bool grow, std::vector<std::vector<int>>* seamsVect)
+{
+	cv::Mat newFrame = image.clone();
+	assert(!newFrame.empty());
+	std::vector<std::vector<int>> vecSeams;
+
+	for(int i = 0; i < seams; i++)
+	{
+		//Gradient Magnitude for intensity of image.
+		cv::Mat gradientMagnitude = computeGradientMagnitude(newFrame);
+		//Use DP to create the real energy map that is used for path calculation.
+		// Strictly using vertical paths for testing simplicity.
+		cv::Mat pathIntensityMat = computePathIntensityMat(gradientMagnitude);
+
+		if(pathIntensityMat.rows == 0 && pathIntensityMat.cols == 0)
+			return false;
+		std::vector<int> seam = getLeastImportantPath(pathIntensityMat);
+		vecSeams.push_back(seam);
+		if(seamsVect)
+			seamsVect->push_back(seam);
+
+		newFrame = removeLeastImportantPath(newFrame, seam);
+
+		if(newFrame.rows == 0 || newFrame.cols == 0)
+			return false;
+	}
+
+	if (grow)
+	{
+		cv::Mat growMat = image.clone();
+
+		for(size_t i = 0; i < vecSeams.size(); i++)
+		{
+			growMat = addLeastImportantPath(growMat,vecSeams[i]);
+		}
+		image = growMat;
+	}
+	else
+	{
+		image = newFrame;
+	}
+	return true;
+}
+
+bool SeamCarving::strechImageVert(cv::Mat& image, int seams, bool grow, std::vector<std::vector<int>>* seamsVect)
+{
+	cv::transpose(image, image);
+	bool ret = strechImage(image, seams, grow, seamsVect);
+	cv::transpose(image, image);
+	return ret;
+}
+
+bool SeamCarving::strechImageWithSeamsImage(cv::Mat& image, cv::Mat& seamsImage, int seams, bool grow)
+{
+	std::vector<std::vector<int>> seamsVect;
+	seamsImage = image.clone();
+
+	bool ret = SeamCarving::strechImage(image, seams, grow, &seamsVect);
+	if(!ret)
+		return false;
+
+	for(size_t i = 0; i < seamsVect.size(); ++i)
+		seamsImage = drawSeam(seamsImage, seamsVect[i]);
+	return true;
+}
+
+cv::Mat SeamCarving::GetEnergyImg(const cv::Mat &img)
+{
+	// find partial derivative of x-axis and y-axis seperately
+	// sum up the partial derivates
+	float pd[] = {1, 2, 1, 0, 0, 0, -1, -2 - 1};
+	cv::Mat xFilter(3, 3, CV_32FC1, pd);
+	cv::Mat yFilter = xFilter.t();
+	cv::Mat grayImg;
+	cv::cvtColor(img, grayImg, cv::COLOR_RGBA2GRAY);
+	cv::Mat dxImg;
+	cv::Mat dyImg;
+
+	cv::filter2D(grayImg, dxImg, 0, xFilter);
+	cv::filter2D(grayImg, dyImg, 0, yFilter);
+	//cv::Mat zeroMat = cv::Mat::zeros(dxImg.rows, dxImg.cols, dxImg.type());
+	//cv::Mat absDxImg;
+	//cv::Mat absDyImg;
+	//cv::absdiff(dxImg, zeroMat, absDxImg);
+	//cv::absdiff(dyImg, zeroMat, absDyImg);
+	cv::Mat absDxImg = cv::abs(dxImg);
+	cv::Mat absDyImg = cv::abs(dyImg);
+
+	cv::Mat energyImg;
+	cv::add(absDxImg, absDyImg, energyImg);
+	return energyImg;
+}
+
+cv::Mat SeamCarving::computeGradientMagnitude(const cv::Mat &frame)
+{
+	cv::Mat grayScale;
+	cv::cvtColor(frame, grayScale, cv::COLOR_RGBA2GRAY);
+	cv::Mat drv = cv::Mat(grayScale.size(), CV_16SC1);
+	cv::Mat drv32f = cv::Mat(grayScale.size(), CV_32FC1);
+	cv::Mat mag = cv::Mat::zeros(grayScale.size(), CV_32FC1);
+	Sobel(grayScale, drv, CV_16SC1, 1, 0);
+	drv.convertTo(drv32f, CV_32FC1);
+	cv::accumulateSquare(drv32f, mag);
+	Sobel(grayScale, drv, CV_16SC1, 0, 1);
+	drv.convertTo(drv32f, CV_32FC1);
+	cv::accumulateSquare(drv32f, mag);
+	cv::sqrt(mag, mag);
+	return mag;
+}
+
+float SeamCarving::intensity(float currIndex, int start, int end)
+{
+	if(start < 0 || start >= end)
+	{
+		return FLT_MAX;
+	}
+	else
+	{
+		return currIndex;
+	}
+}
+
+cv::Mat SeamCarving::computePathIntensityMat(const cv::Mat &rawEnergyMap)
+{
+	cv::Mat pathIntensityMap = cv::Mat(rawEnergyMap.size(), CV_32FC1);
+
+	if(rawEnergyMap.total() == 0 || pathIntensityMap.total() == 0)
+	{
+		return cv::Mat();
+	}
+
+	//First row of intensity paths is the same as the energy map
+	rawEnergyMap.row(0).copyTo(pathIntensityMap.row(0));
+	float max = 0;
+
+	//The rest of them use the DP calculation using the minimum of the 3 pixels above them + their own intensity.
+	for(int row = 1; row < pathIntensityMap.rows; row++)
+	{
+		for(int col = 0; col < pathIntensityMap.cols; col++)
+		{
+			//The initial intensity of the pixel is its raw intensity
+			float pixelIntensity = rawEnergyMap.at<float>(row, col);
+			//The minimum intensity from the current path of the 3 pixels above it is added to its intensity.
+			float p1 = intensity(pathIntensityMap.at<float>(row-1, col-1), col - 1, pathIntensityMap.cols);
+			float p2 = intensity(pathIntensityMap.at<float>(row-1, col), col, pathIntensityMap.cols);
+			float p3 = intensity(pathIntensityMap.at<float>(row-1, col+1), col + 1, pathIntensityMap.cols);
+
+			float minIntensity = std::min(p1, p2);
+			minIntensity = std::min(minIntensity, p3);
+
+			pixelIntensity += minIntensity;
+
+			max = std::max(max, pixelIntensity);
+			pathIntensityMap.at<float>(row, col) = pixelIntensity;
+		}
+	}
+	return pathIntensityMap;
+}
+
+std::vector<int> SeamCarving::getLeastImportantPath(const cv::Mat &importanceMap)
+{
+	if(importanceMap.total() == 0)
+	{
+		return std::vector<int>();
+	}
+
+	//Find the beginning of the least important path. Trying an averaging approach because absolute min wasn't very reliable.
+	float minImportance = importanceMap.at<float>(importanceMap.rows - 1, 0);
+	int minCol = 0;
+	for (int col = 1; col < importanceMap.cols; col++)
+	{
+		float currPixel =importanceMap.at<float>(importanceMap.rows - 1, col);
+		if(currPixel < minImportance)
+		{
+			minCol = col;
+			minImportance = currPixel;
+		}
+	}
+
+	std::vector<int> leastEnergySeam(importanceMap.rows);
+	leastEnergySeam[importanceMap.rows-1] = minCol;
+	for(int row = importanceMap.rows - 2; row >= 0; row--)
+	{
+		float p1 = intensity(importanceMap.at<float>(row, minCol-1), minCol - 1, importanceMap.cols);
+		float p2 = intensity(importanceMap.at<float>(row, minCol), minCol, importanceMap.cols);
+		float p3 = intensity(importanceMap.at<float>(row, minCol+1), minCol + 1, importanceMap.cols);
+		//Adjust the min column for path following
+		if(p1 < p2 && p1 < p3)
+		{
+			minCol -= 1;
+		}
+		else if(p3 < p1 && p3 < p2)
+		{
+			minCol += 1;
+		}
+		leastEnergySeam[row] = minCol;
+	}
+
+	return leastEnergySeam;
+}
+
+cv::Mat SeamCarving::removeLeastImportantPath(const cv::Mat &original, const std::vector<int> &seam)
+{
+	cv::Size orgSize = original.size();
+	// new mat needs to shrink by one collumn
+	cv::Size size = cv::Size(orgSize.width-1, orgSize.height);
+	cv::Mat newMat = cv::Mat(size, original.type());
+
+	for(size_t row = 0; row < seam.size(); row++)
+	{
+		removePixel(original, newMat, row, seam[row]);
+	}
+	return newMat;
+}
+
+void SeamCarving::removePixel(const cv::Mat &original, cv::Mat &outputMat, int row, int minCol)
+{
+	int width = original.cols;
+	int channels = original.channels();
+	int originRowStart = row * channels * width;
+	int newRowStart = row * channels * (width - 1);
+	int firstNum = minCol * channels;
+	unsigned char *rawOrig = original.data;
+	unsigned char *rawOutput = outputMat.data;
+
+	//std::cout << "originRowStart: " << originRowStart << std::endl;
+	//std::cout << "newRowStart: " << newRowStart << std::endl;
+	//std::cout << "firstNum: " << firstNum << std::endl;
+	memcpy(rawOutput + newRowStart, rawOrig + originRowStart, firstNum);
+
+	int originRowMid = originRowStart + (minCol + 1) * channels;
+	int newRowMid = newRowStart + minCol * channels;
+	int secondNum = (width - 1) * channels - firstNum;
+
+	//std::cout << "originRowMid: " << originRowMid << std::endl;
+	//std::cout << "newRowMid: " << newRowMid << std::endl;
+	//std::cout << "secondNum: " << secondNum << std::endl;
+	memcpy(rawOutput + newRowMid, rawOrig + originRowMid, secondNum);
+
+	int leftPixel = minCol - 1;
+	int rightPixel = minCol + 1;
+
+	int byte1 = rawOrig[originRowStart + minCol * channels];
+	int byte2 = rawOrig[originRowStart + minCol * channels + 1];
+	int byte3 = rawOrig[originRowStart + minCol * channels + 2];
+
+	if (rightPixel < width)
+	{
+		int byte1R = rawOrig[originRowStart + rightPixel * channels];
+		int byte2R = rawOrig[originRowStart + rightPixel * channels + 1];
+		int byte3R = rawOrig[originRowStart + rightPixel * channels + 2];
+		rawOutput[newRowStart + minCol * channels] = (unsigned char)((byte1 + byte1R) / 2);
+		rawOutput[newRowStart + minCol * channels + 1] = (unsigned char)((byte2 + byte2R) / 2);
+		rawOutput[newRowStart + minCol * channels + 2] = (unsigned char)((byte3 + byte3R) / 2);
+	}
+
+	if(leftPixel >= 0)
+	{
+		int byte1L = rawOrig[originRowStart + leftPixel*channels];
+		int byte2L = rawOrig[originRowStart + leftPixel*channels+1];
+		int byte3L = rawOrig[originRowStart + leftPixel*channels+2];
+		rawOutput[newRowStart + leftPixel*channels] = (unsigned char) ((byte1 + byte1L)/2);
+		rawOutput[newRowStart + leftPixel*channels+1] = (unsigned char) ((byte2 + byte2L)/2);
+		rawOutput[newRowStart + leftPixel*channels+2] = (unsigned char) ((byte3 + byte3L)/2);
+	}
+}
+
+cv::Mat SeamCarving::addLeastImportantPath(const cv::Mat &original, const std::vector<int> &seam)
+{
+	cv::Size orgSize = original.size();
+	// new mat needs to grow by one column
+	cv::Size size = cv::Size(orgSize.width+1, orgSize.height);
+	cv::Mat newMat = cv::Mat(size, original.type());
+
+	for(size_t row = 0; row < seam.size(); row++)
+	{
+		//std::cout << "row: " << row << ", col: " << seam[row] << std::endl;
+		addPixel(original, newMat, row, seam[row]);
+	}
+	return newMat;
+}
+
+void SeamCarving::addPixel(const cv::Mat &original, cv::Mat &outputMat, int row, int minCol)
+{
+	int width = original.cols;
+	int channels = original.channels();
+	int originRowStart = row * channels * width;
+	int newRowStart = row * channels * (width + 1);
+	int firstNum = (minCol + 1) * channels;
+
+	unsigned char *rawOrig = original.data;
+	unsigned char *rawOutput = outputMat.data;
+
+	memcpy(rawOutput + newRowStart, rawOrig + originRowStart, firstNum);
+
+	memcpy(rawOutput + newRowStart + firstNum, rawOrig + originRowStart + firstNum, channels);
+
+	int originRowMid = originRowStart + ((minCol + 1) * channels);
+	int newRowMid = newRowStart + ((minCol + 2) * channels);
+	int secondNum = (width * channels) - firstNum;
+
+	memcpy(rawOutput + newRowMid, rawOrig + originRowMid, secondNum);
+
+	int leftPixel = minCol - 1;
+	int rightPixel = minCol + 1;
+
+	int byte1 = rawOrig[originRowStart + minCol * channels];
+	int byte2 = rawOrig[originRowStart + minCol * channels + 1];
+	int byte3 = rawOrig[originRowStart + minCol * channels + 2];
+
+	if (rightPixel < width)
+	{
+		int byte1R = rawOrig[originRowStart + rightPixel * channels];
+		int byte2R = rawOrig[originRowStart + rightPixel * channels + 1];
+		int byte3R = rawOrig[originRowStart + rightPixel * channels + 2];
+		rawOutput[newRowStart + minCol * channels] = (unsigned char)((byte1 + byte1R) / 2);
+		rawOutput[newRowStart + minCol * channels + 1] = (unsigned char)((byte2 + byte2R) / 2);
+		rawOutput[newRowStart + minCol * channels + 2] = (unsigned char)((byte3 + byte3R) / 2);
+	}
+
+	if(leftPixel >= 0)
+	{
+		int byte1L = rawOrig[originRowStart + leftPixel*channels];
+		int byte2L = rawOrig[originRowStart + leftPixel*channels+1];
+		int byte3L = rawOrig[originRowStart + leftPixel*channels+2];
+		rawOutput[newRowStart + leftPixel*channels] = (unsigned char) ((byte1 + byte1L)/2);
+		rawOutput[newRowStart + leftPixel*channels+1] = (unsigned char) ((byte2 + byte2L)/2);
+		rawOutput[newRowStart + leftPixel*channels+2] = (unsigned char) ((byte3 + byte3L)/2);
+	}
+}
+
+cv::Mat SeamCarving::drawSeam(const cv::Mat &frame, const std::vector<int> &seam)
+{
+	cv::Mat retMat = frame.clone();
+	for(int row = 0; row < frame.rows; row++)
+	{
+		for(int col = 0; col < frame.cols; col++)
+		{
+			retMat.at<cv::Vec3b>(row, seam[row])[0] = 0;
+			retMat.at<cv::Vec3b>(row, seam[row])[1] = 255;
+			retMat.at<cv::Vec3b>(row, seam[row])[2] = 0;
+		}
+	}
+	return retMat;
+}

SmartCrop/seamcarving.h

@@ -0,0 +1,43 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include <opencv2/core/core.hpp>
+#include <vector>
+
+class SeamCarving
+{
+private:
+	static cv::Mat GetEnergyImg(const cv::Mat &img);
+	static cv::Mat computeGradientMagnitude(const cv::Mat &frame);
+	static float intensity(float currIndex, int start, int end);
+	static cv::Mat computePathIntensityMat(const cv::Mat &rawEnergyMap);
+	static std::vector<int> getLeastImportantPath(const cv::Mat &importanceMap);
+	static cv::Mat removeLeastImportantPath(const cv::Mat &original, const std::vector<int> &seam);
+	static void removePixel(const cv::Mat &original, cv::Mat &outputMap, int row, int minCol);
+	static cv::Mat addLeastImportantPath(const cv::Mat &original, const std::vector<int> &seam);
+	static void addPixel(const cv::Mat &original, cv::Mat &outputMat, int row, int minCol);
+	static cv::Mat drawSeam(const cv::Mat &frame, const std::vector<int> &seam);
+
+public:
+	static bool strechImage(cv::Mat& image, int seams, bool grow, std::vector<std::vector<int>>* seamsVect = nullptr);
+	static bool strechImageVert(cv::Mat& image, int seams, bool grow, std::vector<std::vector<int>>* seamsVect = nullptr);
+	static bool strechImageWithSeamsImage(cv::Mat& image, cv::Mat& seamsImage, int seams, bool grow);
+};

SmartCrop/tokenize.cpp

@@ -0,0 +1,46 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include "tokenize.h"
+
+
+std::vector<std::string> tokenizeBinaryIgnore(const std::string& str, const char delim, const char ignoreBraket, const char escapeChar)
+{
+	std::vector<std::string> tokens;
+	std::string token;
+	bool inBaracket = false;
+	for(size_t i = 0; i < str.size(); ++i)
+	{
+		if(str[i] == delim && !inBaracket && (i == 0 || str[i-1] != escapeChar))
+		{
+			tokens.push_back(token);
+			token.clear();
+		}
+		else
+		{
+			token.push_back(str[i]);
+		}
+		if(ignoreBraket == str[i])
+			inBaracket = !inBaracket;
+	}
+	if(!inBaracket)
+		tokens.push_back(token);
+	return tokens;
+}

SmartCrop/tokenize.h

@@ -0,0 +1,26 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+std::vector<std::string> tokenizeBinaryIgnore(const std::string& str, const char delim, const char ignoreBraket = '\0',
+											  const char escapeChar = '\0');

SmartCrop/utils.cpp

@@ -0,0 +1,80 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include "utils.h"
+
+#include <filesystem>
+#include <vector>
+#include <opencv2/imgproc.hpp>
+
+bool isImagePath(const std::filesystem::path& path)
+{
+	return std::filesystem::is_regular_file(path) && (path.extension() == ".png" || path.extension() == ".jpg" || path.extension() == ".jpeg");
+}
+
+void getImageFiles(const std::filesystem::path& path, std::vector<std::filesystem::path>& paths)
+{
+	if(isImagePath(path))
+	{
+		paths.push_back(path);
+	}
+	else if(std::filesystem::is_directory(path))
+	{
+		for(const std::filesystem::directory_entry& dirent : std::filesystem::directory_iterator(path))
+		{
+			if(std::filesystem::is_directory(dirent.path()))
+				getImageFiles(dirent.path(), paths);
+			else if(isImagePath(dirent.path()))
+				paths.push_back(dirent.path());
+		}
+	}
+}
+
+cv::Rect rectFromPoints(const std::vector<std::pair<cv::Point, int>>& points)
+{
+	int left = std::numeric_limits<int>::max();
+	int right = std::numeric_limits<int>::min();
+	int top = std::numeric_limits<int>::max();
+	int bottom = std::numeric_limits<int>::min();
+
+	for(const std::pair<cv::Point, int>& point : points)
+	{
+		left = point.first.x < left ? point.first.x : left;
+		right = point.first.x > right ? point.first.x : right;
+
+		top = point.first.y < top ? point.first.y : top;
+		bottom = point.first.y > bottom ? point.first.y : bottom;
+	}
+
+	return cv::Rect(left, top, right-left, bottom-top);
+}
+
+double pointDist(const cv::Point2i& pointA, const cv::Point2i& pointB)
+{
+	cv::Vec2i a(pointA.x, pointA.y);
+	cv::Vec2i b(pointB.x, pointB.y);
+	return cv::norm(a-b);
+}
+
+bool pointInRect(const cv::Point2i& point, const cv::Rect& rect)
+{
+	return point.x >= rect.x && point.x <= rect.x+rect.width &&
+		   point.y >= rect.y && point.y <= rect.y+rect.height;
+}

SmartCrop/utils.h

@@ -0,0 +1,34 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include <filesystem>
+#include <vector>
+#include <opencv2/imgproc.hpp>
+
+bool isImagePath(const std::filesystem::path& path);
+
+void getImageFiles(const std::filesystem::path& path, std::vector<std::filesystem::path>& paths);
+
+cv::Rect rectFromPoints(const std::vector<std::pair<cv::Point, int>>& points);
+
+double pointDist(const cv::Point2i& pointA, const cv::Point2i& pointB);
+
+bool pointInRect(const cv::Point2i& point, const cv::Rect& rect);

SmartCrop/yolo.cpp

@@ -0,0 +1,278 @@
+//
+// SmartCrop - A tool for content aware croping of images
+// Copyright (C) 2024 Carl Philipp Klemm
+//
+// This file is part of SmartCrop.
+//
+// SmartCrop is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// SmartCrop is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+#include <opencv2/dnn/dnn.hpp>
+#include <algorithm>
+#include <string>
+#include <stdexcept>
+
+#include "yolo.h"
+#include "readfile.h"
+#include "tokenize.h"
+#include "log.h"
+
+#define INCBIN_PREFIX r
+#include "incbin.h"
+
+INCTXT(defaultClasses, WEIGHT_DIR "/classes.txt");
+INCBIN(defaultModel, WEIGHT_DIR "/yolov8x.onnx");
+
+Yolo::Yolo(const std::filesystem::path &onnxModelPath, const cv::Size &modelInputShape,
+		const std::filesystem::path& classesTxtFilePath, bool runWithOCl)
+{
+	modelPath = onnxModelPath;
+	modelShape = modelInputShape;
+
+	if(classesTxtFilePath.empty())
+	{
+		Log(Log::INFO)<<"Using builtin classes";
+		loadClasses(rdefaultClassesData);
+	}
+	else
+	{
+		std::string classesStr = readFile(classesTxtFilePath);
+		loadClasses(classesStr);
+	}
+
+	if(!modelPath.empty())
+	{
+		net = cv::dnn::readNetFromONNX(modelPath);
+	}
+	else
+	{
+		Log(Log::INFO)<<"Using builtin yolo model";
+		net = cv::dnn::readNetFromONNX((const char*)rdefaultModelData, rdefaultModelSize);
+	}
+	if(runWithOCl)
+	{
+		net.setPreferableBackend(cv::dnn::DNN_BACKEND_DEFAULT);
+		net.setPreferableTarget(cv::dnn::DNN_TARGET_OPENCL);
+	}
+	else
+	{
+		net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
+		net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
+	}
+}
+
+std::vector<Yolo::Detection> Yolo::runInference(const cv::Mat &input)
+{
+	cv::Mat modelInput = input;
+	if (letterBoxForSquare && modelShape.width == modelShape.height)
+		modelInput = formatToSquare(modelInput);
+
+	cv::Mat blob;
+	cv::dnn::blobFromImage(modelInput, blob, 1.0/255.0, modelShape, cv::Scalar(), true, false);
+	net.setInput(blob);
+
+	std::vector<cv::Mat> outputs;
+	net.forward(outputs, net.getUnconnectedOutLayersNames());
+
+	int rows = outputs[0].size[1];
+	int dimensions = outputs[0].size[2];
+
+	bool yolov8 = false;
+	// yolov5 has an output of shape (batchSize, 25200, 85) (Num classes + box[x,y,w,h] + confidence[c])
+	// yolov8 has an output of shape (batchSize, 84,  8400) (Num classes + box[x,y,w,h])
+	if (dimensions > rows) // Check if the shape[2] is more than shape[1] (yolov8)
+	{
+		yolov8 = true;
+		rows = outputs[0].size[2];
+		dimensions = outputs[0].size[1];
+
+		outputs[0] = outputs[0].reshape(1, dimensions);
+		cv::transpose(outputs[0], outputs[0]);
+	}
+	float *data = (float *)outputs[0].data;
+
+	float x_factor = modelInput.cols / modelShape.width;
+	float y_factor = modelInput.rows / modelShape.height;
+
+	std::vector<int> class_ids;
+	std::vector<float> confidences;
+	std::vector<cv::Rect> boxes;
+
+	for (int i = 0; i < rows; ++i)
+	{
+		if (yolov8)
+		{
+			float *classes_scores = data+4;
+
+			cv::Mat scores(1, classes.size(), CV_32FC1, classes_scores);
+			cv::Point class_id;
+			double maxClassScore;
+
+			minMaxLoc(scores, 0, &maxClassScore, 0, &class_id);
+
+			if (maxClassScore > modelScoreThreshold)
+			{
+				confidences.push_back(maxClassScore);
+				class_ids.push_back(class_id.x);
+
+				float x = data[0];
+				float y = data[1];
+				float w = data[2];
+				float h = data[3];
+
+				int left = int((x - 0.5 * w) * x_factor);
+				int top = int((y - 0.5 * h) * y_factor);
+
+				int width = int(w * x_factor);
+				int height = int(h * y_factor);
+
+				boxes.push_back(cv::Rect(left, top, width, height));
+			}
+		}
+		else // yolov5
+		{
+			float confidence = data[4];
+
+			if (confidence >= modelConfidenceThreshold)
+			{
+				float *classes_scores = data+5;
+
+				cv::Mat scores(1, classes.size(), CV_32FC1, classes_scores);
+				cv::Point class_id;
+				double max_class_score;
+
+				minMaxLoc(scores, 0, &max_class_score, 0, &class_id);
+
+				if (max_class_score > modelScoreThreshold)
+				{
+					confidences.push_back(confidence);
+					class_ids.push_back(class_id.x);
+
+					float x = data[0];
+					float y = data[1];
+					float w = data[2];
+					float h = data[3];
+
+					int left = int((x - 0.5 * w) * x_factor);
+					int top = int((y - 0.5 * h) * y_factor);
+
+					int width = int(w * x_factor);
+					int height = int(h * y_factor);
+
+					boxes.push_back(cv::Rect(left, top, width, height));
+				}
+			}
+		}
+
+		data += dimensions;
+	}
+
+	std::vector<int> nms_result;
+	cv::dnn::NMSBoxes(boxes, confidences, modelScoreThreshold, modelNMSThreshold, nms_result);
+
+	std::vector<Yolo::Detection> detections{};
+	for(unsigned long i = 0; i < nms_result.size(); ++i)
+	{
+		int idx = nms_result[i];
+
+		Yolo::Detection result;
+		result.class_id = class_ids[idx];
+		result.confidence = confidences[idx];
+
+		std::random_device rd;
+		std::mt19937 gen(rd());
+		std::uniform_int_distribution<int> dis(100, 255);
+		result.color = cv::Scalar(dis(gen),
+		                          dis(gen),
+		                          dis(gen));
+
+		result.className = classes[result.class_id].first;
+		result.priority = classes[result.class_id].second;
+		clampBox(boxes[idx], input.size());
+		result.box = boxes[idx];
+		detections.push_back(result);
+	}
+
+	return detections;
+}
+
+
+void Yolo::clampBox(cv::Rect& box, const cv::Size& size)
+{
+	if(box.x < 0)
+	{
+		box.width += box.x;
+		box.x = 0;
+	}
+	if(box.y < 0)
+	{
+		box.height += box.y;
+		box.y = 0;
+	}
+	if(box.x+box.width > size.width)
+		box.width = size.width - box.x;
+	if(box.y+box.height > size.height)
+		box.height = size.height - box.y;
+}
+
+void Yolo::loadClasses(const std::string& classesStr)
+{
+	std::vector<std::string> candidateClasses = tokenizeBinaryIgnore(classesStr, '\n', '"', '\\');
+	classes.clear();
+	for(std::string& instance : candidateClasses)
+	{
+		if(instance.size() < 2)
+			continue;
+
+		std::vector<std::string> tokens = tokenizeBinaryIgnore(instance, ',', '"', '\\');
+
+		if(*tokens[0].begin() == '"')
+			instance.erase(tokens[0].begin());
+		if(tokens[0].back() == '"')
+			tokens[0].pop_back();
+		int priority = -1;
+		if(tokens.size() > 1)
+		{
+			try
+			{
+				priority = std::stoi(tokens[1]);
+			}
+			catch(const std::invalid_argument& err)
+			{
+				Log(Log::WARN)<<"unable to get priority for class "<<tokens[0]<<' '<<err.what();
+			}
+		}
+		classes.push_back({tokens[0], priority});
+	}
+}
+
+cv::Mat Yolo::formatToSquare(const cv::Mat &source)
+{
+	int col = source.cols;
+	int row = source.rows;
+	int _max = MAX(col, row);
+	cv::Mat result = cv::Mat::zeros(_max, _max, CV_8UC3);
+	source.copyTo(result(cv::Rect(0, 0, col, row)));
+	return result;
+}
+
+int Yolo::getClassForStr(const std::string& str) const
+{
+	for(size_t i = 0; i < classes.size(); ++i)
+	{
+		if(classes[i].first == str)
+			return i;
+	}
+	return -1;
+}

SmartCrop/yolo.h

@@ -0,0 +1,65 @@
+/* * SmartCrop - A tool for content aware croping of images
+ * Copyright (C) 2024 Carl Philipp Klemm
+ *
+ * This file is part of SmartCrop.
+ *
+ * SmartCrop is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * SmartCrop is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with SmartCrop.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include <fstream>
+#include <vector>
+#include <string>
+#include <random>
+#include <filesystem>
+#include <opencv2/imgproc.hpp>
+#include <opencv2/opencv.hpp>
+#include <opencv2/dnn.hpp>
+
+class Yolo
+{
+public:
+	struct Detection
+	{
+		int class_id = 0;
+		std::string className;
+		float confidence = 0.0;
+		int priority = -1;
+		cv::Scalar color;
+		cv::Rect box;
+	};
+
+private:
+	static constexpr float modelConfidenceThreshold = 0.25;
+	static constexpr float modelScoreThreshold = 0.45;
+	static constexpr float modelNMSThreshold = 0.50;
+
+	std::string modelPath;
+	std::vector<std::pair<std::string, int>> classes;
+	cv::Size2f modelShape;
+	bool letterBoxForSquare = true;
+	cv::dnn::Net net;
+
+	void loadClasses(const std::string& classes);
+	void loadOnnxNetwork(const std::filesystem::path& path);
+	cv::Mat formatToSquare(const cv::Mat &source);
+	static void clampBox(cv::Rect& box, const cv::Size& size);
+
+public:
+	Yolo(const std::filesystem::path &onnxModelPath = "", const cv::Size& modelInputShape = {640, 480},
+		const std::filesystem::path& classesTxtFilePath = "", bool runWithOCl = true);
+	std::vector<Detection> runInference(const cv::Mat &input);
+	int getClassForStr(const std::string& str) const;
+};