yolov5画框重复、大框包小框问题解决，c++、python代码调用onnx

yolov5在训练完成后，获取模型（pt）文件，或者转为onnx文件，对图片进行推理时，会出现以下情况，大框包小框，会导致，明明场景中只有一个目标物而识别出两个或者更多目标物，且画出的框均标记在目标物上，在单张图目标物较多的场景该现象更为严重，具体情况如下图所示。

        如上图所示，右上角帽子的标签就出现了，大框包小框的现象。

        通过查找资料，发现是由于最新的代码，在生成模型，及导出onnx模型时，将anchor box decode过程包含在内。导出的onnx模型如下图所示（仅展示anchor box decode过程）。

用opencv的dnn模块做yolov5目标检测_nihate的博客-CSDN博客_opencv yolov5

        这里有（上面链接），更改代码处理模型的详细过程，总结其中与之相关部分的代码及处理步骤如下所示（或有不同，是在使用时由于报错，对代码做了一点修改）。

models/yolo.py中的detect类中做一下修改（sigmoid可加可不加）

 代码片段（可以复制代码片段）

    def forward(self, x):
        # x = x.copy()  # for profiling
        z = []  # inference output
        # if torch.onnx.is_in_onnx_export():
        for i in range(self.nl):  # 分别对三个输出层处理
            x[i] = self.m[i](x[i])  # conv
            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
            y = x[i].sigmoid()
            z.append(y.view(bs, -1, self.no))
            # z.append(x[i].view(bs, self.na * nx * ny, self.no))
        return torch.cat(z, 1)

在export.py文件中添加 def my_export_onnx 

@try_export
def my_export_onnx(model, im, file, opset, dynamic, prefix=colorstr('ONNX:')):
    print('anchors:', model.yaml['anchors'])
    wtxt = open('class.names', 'w')
    for name in model.names:
        wtxt.write(name+'\n')
    wtxt.close()
    # YOLOv5 ONNX export
    # print(im.shape)
    if not dynamic:
        f = os.path.splitext(file)[0] + '.onnx'
        torch.onnx.export(model, im, f, verbose=False, opset_version=12, input_names=['images'], output_names=['output'])
    else:
        f = os.path.splitext(file)[0] + '_dynamic.onnx'
        torch.onnx.export(model, im, f, verbose=False, opset_version=12, input_names=['images'],
                          output_names=['output'], dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'},  # shape(1,3,640,640)
                                        'output': {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
                                        })
    try:
        import cv2
        net = cv2.dnn.readNet(f)
    except:
        exit(f'export {f} failed')
    exit(f'export {f} sucess')

同时修改export.py文件中的 def export_onnx 。 代码如下所示

@try_export
def export_onnx(model, im, file, opset, dynamic, simplify, prefix=colorstr('ONNX:')):
    # YOLOv5 ONNX export
    check_requirements('onnx')
# ============== 2022.12.14剪枝yolov5的decode部分添加判断代码========================
    my_export_onnx(model, im, file, opset,  False, simplify)
 
    import onnx

    LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...')
    f = file.with_suffix('.onnx')

修改完成后，使用下述命令生成onnx模型文件

python export.py --weights yolov5s.pt --img 640 --batch 1 --include=onnx --simplify 

改过得onnx模型文件，anchor box decode过程如下图所示

 同时，模型推理代码也需要进行更改，python代码如下

import cv2
import argparse
import numpy as np

class yolov5():
    def __init__(self, modelpath, confThreshold=0.5, nmsThreshold=0.5, objThreshold=0.5):
        with open(r'F:\XunLeiDownLoad\yolov5-v6.1-opencv-onnxrun-main\opencv/safetyclass.names', 'rt') as f:
            self.classes = f.read().rstrip('\n').split('\n')
        self.num_classes = len(self.classes)
        if modelpath.endswith('6.onnx'):
            self.inpHeight, self.inpWidth = 1280, 1280
            anchors = [[19, 27, 44, 40, 38, 94], [96, 68, 86, 152, 180, 137], [140, 301, 303, 264, 238, 542],
                       [436, 615, 739, 380, 925, 792]]
            self.stride = np.array([8., 16., 32., 64.])
        else:
            self.inpHeight, self.inpWidth = 640, 640
            anchors = [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]]
            self.stride = np.array([8., 16., 32.])
        self.nl = len(anchors)
        self.na = len(anchors[0]) // 2
        self.grid = [np.zeros(1)] * self.nl
        self.anchor_grid = np.asarray(anchors, dtype=np.float32).reshape(self.nl, -1, 2)
        self.net = cv2.dnn.readNet(modelpath)
        self.confThreshold = confThreshold
        self.nmsThreshold = nmsThreshold
        self.objThreshold = objThreshold
        self._inputNames = ''

    def resize_image(self, srcimg, keep_ratio=True, dynamic=False):
        top, left, newh, neww = 0, 0, self.inpWidth, self.inpHeight
        if keep_ratio and srcimg.shape[0] != srcimg.shape[1]:
            hw_scale = srcimg.shape[0] / srcimg.shape[1]
            if hw_scale > 1:
                newh, neww = self.inpHeight, int(self.inpWidth / hw_scale)
                img = cv2.resize(srcimg, (neww, newh), interpolation=cv2.INTER_AREA)
                if not dynamic:
                    left = int((self.inpWidth - neww) * 0.5)
                    img = cv2.copyMakeBorder(img, 0, 0, left, self.inpWidth - neww - left, cv2.BORDER_CONSTANT,
                                             value=(114, 114, 114))  # add border
            else:
                newh, neww = int(self.inpHeight * hw_scale), self.inpWidth
                img = cv2.resize(srcimg, (neww, newh), interpolation=cv2.INTER_AREA)
                if not dynamic:
                    top = int((self.inpHeight - newh) * 0.5)
                    img = cv2.copyMakeBorder(img, top, self.inpHeight - newh - top, 0, 0, cv2.BORDER_CONSTANT,
                                             value=(114, 114, 114))
        else:
            img = cv2.resize(srcimg, (self.inpWidth, self.inpHeight), interpolation=cv2.INTER_AREA)
        return img, newh, neww, top, left

    def _make_grid(self, nx=20, ny=20):
        xv, yv = np.meshgrid(np.arange(ny), np.arange(nx))
        return np.stack((xv, yv), 2).reshape((-1, 2)).astype(np.float32)

    def preprocess(self, img):
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        img = img.astype(np.float32) / 255.0
        return img

    def postprocess(self, frame, outs, padsize=None):
        frameHeight = frame.shape[0]
        frameWidth = frame.shape[1]
        newh, neww, padh, padw = padsize
        ratioh, ratiow = frameHeight / newh, frameWidth / neww
        # Scan through all the bounding boxes output from the network and keep only the
        # ones with high confidence scores. Assign the box's class label as the class with the highest score.

        confidences = []
        boxes = []
        classIds = []
        for detection in outs:
            if detection[4] > self.objThreshold:
                scores = detection[5:]
                classId = np.argmax(scores)
                confidence = scores[classId] * detection[4]
                if confidence > self.confThreshold:
                    center_x = int((detection[0] - padw) * ratiow)
                    center_y = int((detection[1] - padh) * ratioh)
                    width = int(detection[2] * ratiow)
                    height = int(detection[3] * ratioh)
                    left = int(center_x - width * 0.5)
                    top = int(center_y - height * 0.5)

                    confidences.append(float(confidence))
                    boxes.append([left, top, width, height])
                    classIds.append(classId)
        # Perform non maximum suppression to eliminate redundant overlapping boxes with
        # lower confidences.
        indices = cv2.dnn.NMSBoxes(boxes, confidences, self.confThreshold, self.nmsThreshold).flatten()
        for i in indices:
            box = boxes[i]
            left = box[0]
            top = box[1]
            width = box[2]
            height = box[3]
            frame = self.drawPred(frame, classIds[i], confidences[i], left, top, left + width, top + height)
        return frame

    def drawPred(self, frame, classId, conf, left, top, right, bottom):
        # Draw a bounding box.
        cv2.rectangle(frame, (left, top), (right, bottom), (0, 0, 255), thickness=4)

        label = '%.2f' % conf
        label = '%s:%s' % (self.classes[classId], label)

        # Display the label at the top of the bounding box
        labelSize, baseLine = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
        top = max(top, labelSize[1])
        # cv.rectangle(frame, (left, top - round(1.5 * labelSize[1])), (left + round(1.5 * labelSize[0]), top + baseLine), (255,255,255), cv.FILLED)
        cv2.putText(frame, label, (left, top - 10), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), thickness=2)
        return frame

    def detect(self, srcimg):
        img, newh, neww, padh, padw = self.resize_image(srcimg)
        blob = cv2.dnn.blobFromImage(img, scalefactor=1 / 255.0, swapRB=True)
        # blob = cv2.dnn.blobFromImage(self.preprocess(img))
        # Sets the input to the network
        self.net.setInput(blob, self._inputNames)

        # Runs the forward pass to get output of the output layers
        outs = self.net.forward(self.net.getUnconnectedOutLayersNames())[0].squeeze(axis=0)

        # inference output
        row_ind = 0
        for i in range(self.nl):
            h, w = int(self.inpHeight / self.stride[i]), int(self.inpWidth / self.stride[i])
            length = int(self.na * h * w)
            if self.grid[i].shape[2:4] != (h, w):
                self.grid[i] = self._make_grid(w, h)

            outs[row_ind:row_ind + length, 0:2] = (outs[row_ind:row_ind + length, 0:2] * 2. - 0.5 + np.tile(
                self.grid[i], (self.na, 1))) * int(self.stride[i])
            outs[row_ind:row_ind + length, 2:4] = (outs[row_ind:row_ind + length, 2:4] * 2) ** 2 * np.repeat(
                self.anchor_grid[i], h * w, axis=0)
            row_ind += length
        srcimg = self.postprocess(srcimg, outs, padsize=(newh, neww, padh, padw))
        return srcimg

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--imgpath', type=str, default=r'E:\code\detect\yolov5\dataset\safety_clothing\test_safety_clothing\images1\398404624-1-16_6675.jpg', help="image path")
    parser.add_argument('--modelpath', type=str, default=r'E:\code\detect\yolov5\runs\train\all_safetly\best_all_che.onnx')
    parser.add_argument('--confThreshold', default=0.3, type=float, help='class confidence')
    parser.add_argument('--nmsThreshold', default=0.5, type=float, help='nms iou thresh')
    parser.add_argument('--objThreshold', default=0.3, type=float, help='object confidence')
    args = parser.parse_args()

    yolonet = yolov5(args.modelpath, confThreshold=args.confThreshold, nmsThreshold=args.nmsThreshold,
                     objThreshold=args.objThreshold)
    srcimg = cv2.imread(args.imgpath)
    srcimg = yolonet.detect(srcimg)

    winName = 'Deep learning object detection in OpenCV'
    cv2.namedWindow(winName, 0)
    cv2.imshow(winName, srcimg)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

safetyclass.names 为标签文件，下图是截图展示

c++推理代码如下

#include 
#include 
#include 
#include 
#include 
#include 

using namespace cv;
using namespace dnn;
using namespace std;

struct Net_config
{
	float confThreshold; // Confidence threshold
	float nmsThreshold;  // Non-maximum suppression threshold
	float objThreshold;  //Object Confidence threshold
	string modelpath;
};

//int endsWith(const string& s, const string& sub) {
//	return s.rfind(sub) == (s.length() - sub.length()) ? 1 : 0;
//}

const float anchors_640[3][6] = { {10.0,  13.0, 16.0,  30.0,  33.0,  23.0},
								 {30.0,  61.0, 62.0,  45.0,  59.0,  119.0},
								 {116.0, 90.0, 156.0, 198.0, 373.0, 326.0} };

//const float anchors_1280[4][6] = { {19, 27, 44, 40, 38, 94},{96, 68, 86, 152, 180, 137},{140, 301, 303, 264, 238, 542},
//					   {436, 615, 739, 380, 925, 792} };

class YOLO
{
public:
	explicit YOLO(const Net_config& config);
	void detect(Mat& frame);
private:
	float* anchors;
	int num_stride;
	int inpWidth;
	int inpHeight;
	vector class_names;
//	int num_class;
	
	float confThreshold;
	float nmsThreshold;
	float objThreshold;
	const bool keep_ratio = true;
	Net net;
	void drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid);
	Mat resize_image(const Mat& srcimg, int *newh, int *neww, int *top, int *left) const;
};

YOLO::YOLO(const Net_config& config)
{
	this->confThreshold = config.confThreshold;
	this->nmsThreshold = config.nmsThreshold;
	this->objThreshold = config.objThreshold;

	this->net = readNet(config.modelpath);
//	ifstream ifs("F:\\XunLeiDownLoad\\yolov5-v6.1-opencv-onnxrun-main\\opencv/class.names");
    ifstream ifs(R"(F:\XunLeiDownLoad\yolov5-v6.1-opencv-onnxrun-main\opencv/safetyclass.names)");
	string line;
	while (getline(ifs, line)) this->class_names.push_back(line);
//	this->num_class = class_names.size();

//	if (endsWith(config.modelpath, "6.onnx"))
//	{
//		anchors = (float*)anchors_1280;
//		this->num_stride = 4;
//		this->inpHeight = 1280;
//		this->inpWidth = 1280;
//	}
//	else
//	{
		anchors = (float*)anchors_640;
		this->num_stride = 3;
		this->inpHeight = 640;
		this->inpWidth = 640;
//	}
}

Mat YOLO::resize_image(const Mat& srcimg, int *newh, int *neww, int *top, int *left) const
{
	int srch = srcimg.rows, srcw = srcimg.cols;
	*newh = this->inpHeight;
	*neww = this->inpWidth;
	Mat dstimg;
	if (this->keep_ratio && srch != srcw) {
		float hw_scale = (float)srch / srcw;
		if (hw_scale > 1) {
			*newh = this->inpHeight;
			*neww = int(this->inpWidth / hw_scale);
			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
			*left = int((this->inpWidth - *neww) * 0.5);
			copyMakeBorder(dstimg, dstimg, 0, 0, *left, this->inpWidth - *neww - *left, BORDER_CONSTANT, 114);
		}
		else {
			*newh = (int)this->inpHeight * hw_scale;
			*neww = this->inpWidth;
			resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
			*top = (int)(this->inpHeight - *newh) * 0.5;
			copyMakeBorder(dstimg, dstimg, *top, this->inpHeight - *newh - *top, 0, 0, BORDER_CONSTANT, 114);
		}
	}
	else {
		resize(srcimg, dstimg, Size(*neww, *newh), INTER_AREA);
	}
	return dstimg;
}

void YOLO::drawPred(float conf, int left, int top, int right, int bottom, Mat& frame, int classid)   // Draw the predicted bounding box
{
	//Draw a rectangle displaying the bounding box
	rectangle(frame, Point(left, top), Point(right, bottom), Scalar(0, 0, 255), 2);

	//Get the label for the class name and its confidence
	string label = format("%.2f", conf);
	label = this->class_names[classid] + ":" + label;

	//Display the label at the top of the bounding box
	int baseLine;
	Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
	top = max(top, labelSize.height);
	//rectangle(frame, Point(left, top - int(1.5 * labelSize.height)), Point(left + int(1.5 * labelSize.width), top + baseLine), Scalar(0, 255, 0), FILLED);
	putText(frame, label, Point(left, top), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(0, 255, 0), 1);
}

void YOLO::detect(Mat& frame)
{
	int newh = 0, neww = 0, padh = 0, padw = 0;
	Mat dstimg = this->resize_image(frame, &newh, &neww, &padh, &padw);
	Mat blob = blobFromImage(dstimg, 1 / 255.0, Size(this->inpWidth, this->inpHeight), Scalar(0, 0, 0), true, false);
	this->net.setInput(blob);
	vector outs;
	this->net.forward(outs, this->net.getUnconnectedOutLayersNames());

	int num_proposal = outs[0].size[1];
	int nout = outs[0].size[2];
	if (outs[0].dims > 2)
	{
		outs[0] = outs[0].reshape(0, num_proposal);
	}
	/generate proposals
	vector confidences;
	vector boxes;
	vector classIds;
	float ratioh = (float)frame.rows / newh, ratiow = (float)frame.cols / neww;
	int n = 0, q = 0, i = 0, j = 0, row_ind = 0; ///xmin,ymin,xamx,ymax,box_score,class_score
	auto* pdata = (float*)outs[0].data;
	for (n = 0; n < this->num_stride; n++)   ///特征图尺度
	{
		const float stride = pow(2, n + 3);
		int num_grid_x = (int)ceil((this->inpWidth / stride));
		int num_grid_y = (int)ceil((this->inpHeight / stride));
		for (q = 0; q < 3; q++)    ///anchor
		{
			const float anchor_w = this->anchors[n * 6 + q * 2];
			const float anchor_h = this->anchors[n * 6 + q * 2 + 1];
			for (i = 0; i < num_grid_y; i++)
			{
				for (j = 0; j < num_grid_x; j++)
				{
					float box_score = pdata[4];
					if (box_score > this->objThreshold)
					{
						Mat scores = outs[0].row(row_ind).colRange(5, nout);
						Point classIdPoint;
						double max_class_socre;
						// Get the value and location of the maximum score
						minMaxLoc(scores, 0, &max_class_socre, 0, &classIdPoint);
						max_class_socre *= box_score;
						if (max_class_socre > this->confThreshold)
						{ 
							const int class_idx = classIdPoint.x;
							float cx = (pdata[0] * 2.f - 0.5f + j) * stride;  ///cx
							float cy = (pdata[1] * 2.f - 0.5f + i) * stride;   ///cy
							float w = powf(pdata[2] * 2.f, 2.f) * anchor_w;   ///w
							float h = powf(pdata[3] * 2.f, 2.f) * anchor_h;  ///h

							int left = int((cx - padw - 0.5 * w)*ratiow);
							int top = int((cy - padh - 0.5 * h)*ratioh);

							confidences.push_back((float)max_class_socre);
//							boxes.push_back(Rect(left, top, (int)(w*ratiow), (int)(h*ratioh)));
                            boxes.emplace_back(left, top, (int)(w*ratiow), (int)(h*ratioh));
							classIds.push_back(class_idx);
						}
					}
					row_ind++;
					pdata += nout;
				}
			}
		}
	}

	// Perform non maximum suppression to eliminate redundant overlapping boxes with
	// lower confidences
	vector indices;
    /*dnn::NMSBoxes
     * 作用：根据给定的检测boxes和对应的scores进行NMS（非极大值抑制）处理
     * NMSBoxes(bboxes,
             scores,
             score_threshold,
             nms_threshold,
             eta=None,
             top_k=None)
            参数：
            boxes: 待处理的边界框 bounding boxes
            scores: 对于于待处理边界框的 scores
            score_threshold: 用于过滤 boxes 的 score 阈值
            nms_threshold: NMS 用到的阈值
            indices: NMS 处理后所保留的边界框的索引值
            eta: 自适应阈值公式中的相关系数：
     * */
	dnn::NMSBoxes(boxes, confidences, this->confThreshold, this->nmsThreshold, indices);
//	for (size_t i = 0; i < indices.size(); ++i)
//	{
//		int idx = indices[i];
//		Rect box = boxes[idx];
//		this->drawPred(confidences[idx], box.x, box.y,
//			box.x + box.width, box.y + box.height, frame, classIds[idx]);
//	}
    for (int idx : indices)
    {
        Rect box = boxes[idx];
        this->drawPred(confidences[idx], box.x, box.y,
                       box.x + box.width, box.y + box.height, frame, classIds[idx]);
    }
}

int main()
{
//	Net_config yolo_nets = { 0.3, 0.5, 0.3, "F:\\XunLeiDownLoad\\yolov5-v6.1-opencv-onnxrun-main\\opencv/weights/yolov5s.onnx" };
  Net_config yolo_nets = { 0.3, 0.5, 0.3, R"(E:\code\detect\yolov5\runs\train\all_safetly\best_all_che.onnx)" };
	YOLO yolo_model(yolo_nets);
	string imgpath = R"(E:\code\detect\yolov5\dataset\safety_clothing\test_safety_clothing\images1\398404624-1-16_6675.jpg)";
	Mat srcimg = imread(imgpath);
	yolo_model.detect(srcimg);
    string saveimg_path= R"(E:\code\detect\yolov5\testsave\safety2.jpg)";
    imwrite(saveimg_path, srcimg);


//	static const string kWinName = "Deep learning object detection in OpenCV";
//	namedWindow(kWinName, WINDOW_NORMAL);
//	imshow(kWinName, srcimg);
//	waitKey(5000);
//	destroyAllWindows();
}

经过上述过程后图片的推理结果如下所示

 

 可以看出，大框包小框的现象消失。

参考文章：
用opencv的dnn模块做yolov5目标检测_nihate的博客-CSDN博客_opencv yolov5windows下最新yolov5转ncnn教程(支持u版yolov5(ultralytics版)v5.0)_五四三两幺-发射!的博客-CSDN博客_yolov5转换xnno

原代码地址：https://github.com/hpc203/yolov5-v6.1-opencv-onnxrun