devbox
AllinToyou
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

Opencv C++实现yolov5部署onnx模型完成目标检测_c++ opencv yolo onnx

作者：AllinToyou | 2024-03-24 06:06:19

赞

踩

c++ opencv yolo onnx

代码分析：

头文件
  1. #include <fstream>  //文件
  2. #include <sstream>  //流
  3. #include <iostream>
  4. #include <opencv2/dnn.hpp>        //深度学习模块-仅提供推理功能
  5. #include <opencv2/imgproc.hpp>    //图像处理模块
  6. #include <opencv2/highgui.hpp>    //媒体的输入输出/视频捕捉/图像和视频的编码解码/图形界面的接口
命名空间
  1. using namespace cv;
  2. using namespace dnn;
  3. using namespace std;
结构体 Net_config
  1. struct Net_config{
  2. 	float confThreshold; // 置信度阈值
  3. 	float nmsThreshold;  // 非最大抑制阈值
  4. 	float objThreshold;  // 对象置信度阈值
  5. 	string modelpath;
  6. };

里面存了三个阈值和模型地址，其中置信度，顾名思义，看检测出来的物体的精准度。以测量值为中心，在一定范围内，真值出现在该范围内的几率。

endsWith()函数 判断sub是不是s的子串
  1. int endsWith(string s, string sub) {
  2. 	return s.rfind(sub) == (s.length() - sub.length()) ? 1 : 0;
  3. }
anchors_640图像接收数组
  1. const float anchors_640[3][6] = { {10.0,  13.0, 16.0,  30.0,  33.0,  23.0},
  2. 				  {30.0,  61.0, 62.0,  45.0,  59.0,  119.0},
  3. 				  {116.0, 90.0, 156.0, 198.0, 373.0, 326.0} };

根据图像大小，选择相应长度的二维数组。深度为3，每层6个位置。

YOLO类
  1. class YOLO{
  2. public:
  3. 	YOLO(Net_config config); //构造函数
  4. 	void detect(Mat& frame); //通过图像参数，进行目标检测
  5. private:
  6. 	float* anchors;
  7. 	int num_stride;
  8. 	int inpWidth;
  9. 	int inpHeight;
  10. 	vector<string> class_names;
  11. 	int num_class;
  12. 	float confThreshold;
  13. 	float nmsThreshold;
  14. 	float objThreshold;
  15. 	const bool keep_ratio = true;
  16. 	Net net;
  17. 	void drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid);
  18. 	Mat resize_image(Mat srcimg, int* newh, int* neww, int* top, int* left);
  19. };
YOLO类构造函数的重载
  1. YOLO::YOLO(Net_config config){
  2. 	this->confThreshold = config.confThreshold;
  3. 	this->nmsThreshold = config.nmsThreshold;
  4. 	this->objThreshold = config.objThreshold;
  5. 	this->net = readNet(config.modelpath);
  6. 	ifstream ifs("class.names"); //class.name中写入标签内容，当前只有'person'，位置与当前.cpp文件同级
  7. 	string line;
  8. 	while (getline(ifs, line)) this->class_names.push_back(line);
  9. 	this->num_class = class_names.size();
  10. 	if (endsWith(config.modelpath, "6.onnx")){ //根据onnx的输入图像格式 选择分辨率 当前为1280x1280 可灵活调整
  11. 		anchors = (float*)anchors_1280;
  12. 		this->num_stride = 4;      //深度
  13. 		this->inpHeight = 1280;    //高
  14. 		this->inpWidth = 1280;     //宽
  15. 	}
  16. 	else{                                       //当前为640x640 可以resize满足onnx需求 也可以调整数组或if中的onnx判断
  17. 		anchors = (float*)anchors_640;
  18. 		this->num_stride = 3;
  19. 		this->inpHeight = 640;
  20. 		this->inpWidth = 640;
  21. 	}
  22. }

重新规定图像大小函数 resize_image()
  1. Mat YOLO::resize_image(Mat srcimg, int* newh, int* neww, int* top, int* left){//传入图像以及需要改变的参数
  2. 	int srch = srcimg.rows, srcw = srcimg.cols;
  3. 	*newh = this->inpHeight;
  4. 	*neww = this->inpWidth;
  5. 	Mat dstimg;
  6. 	if (this->keep_ratio && srch != srcw) {
  7. 		float hw_scale = (float)srch / srcw;
  8. 		if (hw_scale > 1) {
  9. 			*newh = this->inpHeight;
  10. 			*neww = int(this->inpWidth / hw_scale);
  11. 			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  12. 			*left = int((this->inpWidth - *neww) * 0.5);
  13. 			copyMakeBorder(dstimg, dstimg, 0, 0, *left, this->inpWidth - *neww - *left, BORDER_CONSTANT, 114);
  14. 		}else {
  15. 			*newh = (int)this->inpHeight * hw_scale;
  16. 			*neww = this->inpWidth;
  17. 			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  18. 			*top = (int)(this->inpHeight - *newh) * 0.5;
  19. 			copyMakeBorder(dstimg, dstimg, *top, this->inpHeight - *newh - *top, 0, 0, BORDER_CONSTANT, 114);
  20. 		}
  21. 	}else {
  22. 		resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  23. 	}
  24. 	return dstimg;
  25. }

绘制预测的边界框函数 drawPred()
  1. void YOLO::drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid){
  2. 	//绘制一个显示边界框的矩形
  3. 	rectangle(frame, Point(left, top), Point(right, bottom), Scalar(0, 0, 255), 2);
  4.  
  5. 	//获取类名的标签及其置信度
  6. 	string label = format("%.2f", conf);
  7. 	label = this->class_names[classid] + ":" + label;
  8.  
  9. 	//在边界框顶部显示标签
  10. 	int baseLine;
  11. 	Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
  12. 	top = max(top, labelSize.height);
  13. 	//rectangle(frame, Point(left, top - int(1.5 * labelSize.height)), Point(left + int(1.5 * labelSize.width), top + baseLine), Scalar(0, 255, 0), FILLED);
  14. 	putText(frame, label, Point(left, top), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(0, 255, 0), 1);
  15. }

【核心代码】检测函数 detect()
  1. void YOLO::detect(Mat& frame){
  2. 	int newh = 0, neww = 0, padh = 0, padw = 0;
  3. 	Mat dstimg = this->resize_image(frame, &newh, &neww, &padh, &padw);
  4. 	Mat blob = blobFromImage(dstimg, 1 / 255.0, Size(this->inpWidth, this->inpHeight), Scalar(0, 0, 0), true, false);
  5. 	this->net.setInput(blob);
  6. 	vector<Mat> outs;
  7. 	this->net.forward(outs, this->net.getUnconnectedOutLayersNames());
  8. 	int num_proposal = outs[0].size[1];
  9. 	int nout = outs[0].size[2];
  10. 	if (outs[0].dims > 2){
  11. 		outs[0] = outs[0].reshape(0, num_proposal);
  12. 	}
  13. 	
  14. 	vector<float> confidences;
  15. 	vector<Rect> boxes;
  16. 	vector<int> classIds;
  17. 	float ratioh = (float)frame.rows / newh, ratiow = (float)frame.cols / neww;
  18. 	int n = 0, q = 0, i = 0, j = 0, row_ind = 0;  //xmin,ymin,xamx,ymax,box_score,class_score
  19. 	float* pdata = (float*)outs[0].data;
  20. 	for (n = 0; n < this->num_stride; n++){ //特征图尺度
  21. 		const float stride = pow(2, n + 3);
  22. 		int num_grid_x = (int)ceil((this->inpWidth / stride));
  23. 		int num_grid_y = (int)ceil((this->inpHeight / stride));
  24. 		for (q = 0; q < 3; q++){
  25. 			const float anchor_w = this->anchors[n * 6 + q * 2];
  26. 			const float anchor_h = this->anchors[n * 6 + q * 2 + 1];
  27. 			for (i = 0; i < num_grid_y; i++){
  28. 				for (j = 0; j < num_grid_x; j++){
  29. 					float box_score = pdata[4];
  30. 					if (box_score > this->objThreshold){
  31. 						Mat scores = outs[0].row(row_ind).colRange(5, nout);
  32. 						Point classIdPoint;
  33. 						double max_class_socre;
  34. 						//获取最高分的值和位置
  35. 						minMaxLoc(scores, 0, &max_class_socre, 0, &classIdPoint);
  36. 						max_class_socre *= box_score;
  37. 						if (max_class_socre > this->confThreshold){
  38. 							const int class_idx = classIdPoint.x;
  39. 							float cx = pdata[0];  //cx
  40. 							float cy = pdata[1];  //cy
  41. 							float w = pdata[2];   //w
  42. 							float h = pdata[3];   //h
  43. 							int left = int((cx - padw - 0.5 * w) * ratiow);
  44. 							int top = int((cy - padh - 0.5 * h) * ratioh);
  45. 							confidences.push_back((float)max_class_socre);
  46. 							boxes.push_back(Rect(left, top, (int)(w * ratiow), (int)(h * ratioh)));
  47. 							classIds.push_back(class_idx);
  48. 						}
  49. 					}
  50. 					row_ind++;
  51. 					pdata += nout;
  52. 				}
  53. 			}
  54. 		}
  55. 	}
  56. 	// 执行非最大抑制以消除冗余重叠框
  57. 	// 置信度较低
  58. 	vector<int> indices;
  59. 	dnn::NMSBoxes(boxes, confidences, this->confThreshold, this->nmsThreshold, indices);
  60. 	for (size_t i = 0; i < indices.size(); ++i){
  61. 		int idx = indices[i];
  62. 		Rect box = boxes[idx];
  63. 		this->drawPred(confidences[idx], box.x, box.y,box.x + box.width, box.y + box.height, frame, classIds[idx]);
  64. 	}
  65. }

主函数

  1. int main(){
  2.         //加载onnx模型
  3. 	Net_config yolo_nets = { 0.3, 0.5, 0.3, "yolov5n_person_320.onnx" };
  4. 	YOLO yolo_model(yolo_nets);
  5.         //加载单张图片
  6. 	string imgpath = "112.png";
  7. 	Mat srcimg = imread(imgpath);
  8. 	//开始检测
  9.         yolo_model.detect(srcimg);
  10. 	static const string kWinName = "Deep learning object detection in OpenCV";
  11. 	namedWindow(kWinName, WINDOW_NORMAL);
  12. 	imshow(kWinName, srcimg); //显示图片
  13. 	waitKey(0);               //保持停留
  14. 	destroyAllWindows();      //关闭窗口并取消分配任何相关的内存使用
  15. }

完整代码

  1. #include <fstream>
  2. #include <sstream>
  3. #include <iostream>
  4. #include <opencv2/dnn.hpp>
  5. #include <opencv2/imgproc.hpp>
  6. #include <opencv2/highgui.hpp>
  7.  
  8. using namespace cv;
  9. using namespace dnn;
  10. using namespace std;
  11.  
  12. struct Net_config
  13. {
  14. 	float confThreshold; // Confidence threshold
  15. 	float nmsThreshold;  // Non-maximum suppression threshold
  16. 	float objThreshold;  //Object Confidence threshold
  17. 	string modelpath;
  18. };
  19.  
  20. int endsWith(string s, string sub) {
  21. 	return s.rfind(sub) == (s.length() - sub.length()) ? 1 : 0;
  22. }
  23.  
  24. const float anchors_640[3][6] = { {10.0,  13.0, 16.0,  30.0,  33.0,  23.0},
  25. 				  {30.0,  61.0, 62.0,  45.0,  59.0,  119.0},
  26. 				  {116.0, 90.0, 156.0, 198.0, 373.0, 326.0} };
  27.  
  28. const float anchors_1280[4][6] = { {19, 27, 44, 40, 38, 94},
  29.                                    {96, 68, 86, 152, 180, 137},
  30.                                    {140, 301, 303, 264, 238, 542},
  31. 				   {436, 615, 739, 380, 925, 792} };
  32.  
  33. class YOLO
  34. {
  35. public:
  36. 	YOLO(Net_config config);
  37. 	void detect(Mat& frame);
  38. private:
  39. 	float* anchors;
  40. 	int num_stride;
  41. 	int inpWidth;
  42. 	int inpHeight;
  43. 	vector<string> class_names;
  44. 	int num_class;
  45.  
  46. 	float confThreshold;
  47. 	float nmsThreshold;
  48. 	float objThreshold;
  49. 	const bool keep_ratio = true;
  50. 	Net net;
  51. 	void drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid);
  52. 	Mat resize_image(Mat srcimg, int* newh, int* neww, int* top, int* left);
  53. };
  54.  
  55. YOLO::YOLO(Net_config config)
  56. {
  57. 	this->confThreshold = config.confThreshold;
  58. 	this->nmsThreshold = config.nmsThreshold;
  59. 	this->objThreshold = config.objThreshold;
  60.  
  61. 	this->net = readNet(config.modelpath);
  62. 	ifstream ifs("class.names");
  63. 	string line;
  64. 	while (getline(ifs, line)) this->class_names.push_back(line);
  65. 	this->num_class = class_names.size();
  66.  
  67. 	if (endsWith(config.modelpath, "6.onnx"))
  68. 	{
  69. 		anchors = (float*)anchors_1280;
  70. 		this->num_stride = 4;
  71. 		this->inpHeight = 1280;
  72. 		this->inpWidth = 1280;
  73. 	}
  74. 	else
  75. 	{
  76. 		anchors = (float*)anchors_640;
  77. 		this->num_stride = 3;
  78. 		this->inpHeight = 640;
  79. 		this->inpWidth = 640;
  80. 	}
  81. }
  82.  
  83. Mat YOLO::resize_image(Mat srcimg, int* newh, int* neww, int* top, int* left)
  84. {
  85. 	int srch = srcimg.rows, srcw = srcimg.cols;
  86. 	*newh = this->inpHeight;
  87. 	*neww = this->inpWidth;
  88. 	Mat dstimg;
  89. 	if (this->keep_ratio && srch != srcw) {
  90. 		float hw_scale = (float)srch / srcw;
  91. 		if (hw_scale > 1) {
  92. 			*newh = this->inpHeight;
  93. 			*neww = int(this->inpWidth / hw_scale);
  94. 			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  95. 			*left = int((this->inpWidth - *neww) * 0.5);
  96. 			copyMakeBorder(dstimg, dstimg, 0, 0, *left, this->inpWidth - *neww - *left, BORDER_CONSTANT, 114);
  97. 		}
  98. 		else {
  99. 			*newh = (int)this->inpHeight * hw_scale;
  100. 			*neww = this->inpWidth;
  101. 			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  102. 			*top = (int)(this->inpHeight - *newh) * 0.5;
  103. 			copyMakeBorder(dstimg, dstimg, *top, this->inpHeight - *newh - *top, 0, 0, BORDER_CONSTANT, 114);
  104. 		}
  105. 	}
  106. 	else {
  107. 		resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
  108. 	}
  109. 	return dstimg;
  110. }
  111.  
  112. void YOLO::drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid)   // Draw the predicted bounding box
  113. {
  114. 	//Draw a rectangle displaying the bounding box
  115. 	rectangle(frame, Point(left, top), Point(right, bottom), Scalar(0, 0, 255), 2);
  116.  
  117. 	//Get the label for the class name and its confidence
  118. 	string label = format("%.2f", conf);
  119. 	label = this->class_names[classid] + ":" + label;
  120.  
  121. 	//Display the label at the top of the bounding box
  122. 	int baseLine;
  123. 	Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
  124. 	top = max(top, labelSize.height);
  125. 	//rectangle(frame, Point(left, top - int(1.5 * labelSize.height)), Point(left + int(1.5 * labelSize.width), top + baseLine), Scalar(0, 255, 0), FILLED);
  126. 	putText(frame, label, Point(left, top), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(0, 255, 0), 1);
  127. }
  128.  
  129. void YOLO::detect(Mat& frame)
  130. {
  131. 	int newh = 0, neww = 0, padh = 0, padw = 0;
  132. 	Mat dstimg = this->resize_image(frame, &newh, &neww, &padh, &padw);
  133. 	Mat blob = blobFromImage(dstimg, 1 / 255.0, Size(this->inpWidth, this->inpHeight), Scalar(0, 0, 0), true, false);
  134. 	this->net.setInput(blob);
  135. 	vector<Mat> outs;
  136. 	this->net.forward(outs, this->net.getUnconnectedOutLayersNames());
  137.  
  138. 	int num_proposal = outs[0].size[1];
  139. 	int nout = outs[0].size[2];
  140. 	if (outs[0].dims > 2)
  141. 	{
  142. 		outs[0] = outs[0].reshape(0, num_proposal);
  143. 	}
  144. 	/generate proposals
  145. 	vector<float> confidences;
  146. 	vector<Rect> boxes;
  147. 	vector<int> classIds;
  148. 	float ratioh = (float)frame.rows / newh, ratiow = (float)frame.cols / neww;
  149. 	int n = 0, q = 0, i = 0, j = 0, row_ind = 0; ///xmin,ymin,xamx,ymax,box_score,class_score
  150. 	float* pdata = (float*)outs[0].data;
  151. 	for (n = 0; n < this->num_stride; n++)   ///特征图尺度
  152. 	{
  153. 		const float stride = pow(2, n + 3);
  154. 		int num_grid_x = (int)ceil((this->inpWidth / stride));
  155. 		int num_grid_y = (int)ceil((this->inpHeight / stride));
  156. 		for (q = 0; q < 3; q++)    ///anchor
  157. 		{
  158. 			const float anchor_w = this->anchors[n * 6 + q * 2];
  159. 			const float anchor_h = this->anchors[n * 6 + q * 2 + 1];
  160. 			for (i = 0; i < num_grid_y; i++)
  161. 			{
  162. 				for (j = 0; j < num_grid_x; j++)
  163. 				{
  164. 					float box_score = pdata[4];
  165. 					if (box_score > this->objThreshold)
  166. 					{
  167. 						Mat scores = outs[0].row(row_ind).colRange(5, nout);
  168. 						Point classIdPoint;
  169. 						double max_class_socre;
  170. 						// Get the value and location of the maximum score
  171. 						minMaxLoc(scores, 0, &max_class_socre, 0, &classIdPoint);
  172. 						max_class_socre *= box_score;
  173. 						if (max_class_socre > this->confThreshold)
  174. 						{
  175. 							const int class_idx = classIdPoint.x;
  176. 							//float cx = (pdata[0] * 2.f - 0.5f + j) * stride;  ///cx
  177. 							//float cy = (pdata[1] * 2.f - 0.5f + i) * stride;   ///cy
  178. 							//float w = powf(pdata[2] * 2.f, 2.f) * anchor_w;   ///w
  179. 							//float h = powf(pdata[3] * 2.f, 2.f) * anchor_h;  ///h
  180.  
  181. 							float cx = pdata[0];  ///cx
  182. 							float cy = pdata[1];   ///cy
  183. 							float w = pdata[2];   ///w
  184. 							float h = pdata[3];  ///h
  185.  
  186. 							int left = int((cx - padw - 0.5 * w) * ratiow);
  187. 							int top = int((cy - padh - 0.5 * h) * ratioh);
  188.  
  189. 							confidences.push_back((float)max_class_socre);
  190. 							boxes.push_back(Rect(left, top, (int)(w * ratiow), (int)(h * ratioh)));
  191. 							classIds.push_back(class_idx);
  192. 						}
  193. 					}
  194. 					row_ind++;
  195. 					pdata += nout;
  196. 				}
  197. 			}
  198. 		}
  199. 	}
  200.  
  201. 	// Perform non maximum suppression to eliminate redundant overlapping boxes with
  202. 	// lower confidences
  203. 	vector<int> indices;
  204. 	dnn::NMSBoxes(boxes, confidences, this->confThreshold, this->nmsThreshold, indices);
  205. 	for (size_t i = 0; i < indices.size(); ++i)
  206. 	{
  207. 		int idx = indices[i];
  208. 		Rect box = boxes[idx];
  209. 		this->drawPred(confidences[idx], box.x, box.y,
  210. 			box.x + box.width, box.y + box.height, frame, classIds[idx]);
  211. 	}
  212. }
  213.  
  214. int main()
  215. {
  216. 	Net_config yolo_nets = { 0.3, 0.5, 0.3, "yolov5n_person_320.onnx" };
  217. 	YOLO yolo_model(yolo_nets);
  218.  
  219. 	//string imgpath = "112.png";
  220. 	//Mat srcimg = imread(imgpath);
  221. 	//yolo_model.detect(srcimg);
  222.         
  223.         int n = 588;
  224. 	for (int i = 1; i <= n; i++) {
  225. 		string s = to_string(i) + ".png";
  226. 		string imgpath = "F://test//p1//yanfa2//bh//cc//" + s;
  227. 		cout << imgpath << endl;
  228.  
  229. 		Mat srcimg = imread(imgpath);
  230. 		yolo_model.detect(srcimg);
  231. 		imwrite("F://test//p2//yanfa2//bh//cc//" + s, srcimg);
  232. 	}
  233.  
  234. 	//static const string kWinName = "Deep learning object detection in OpenCV";
  235. 	//namedWindow(kWinName, WINDOW_NORMAL);
  236. 	//imshow(kWinName, srcimg);
  237. 	//waitKey(0);
  238. 	//destroyAllWindows();
  239. }

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/AllinToyou/article/detail/300348
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号