YOLO-V3实时检测实现（opencv+python实现）

YOLO-V3实时检测实现（opencv+python实现）_Keep_Trying_Go的博客-CSDN博客_opencv yolov3

import os
import cv2
import numpy as np
import matplotlib.pyplot as plt
import time
# 读取网络配置文件和权重文件
net = cv2.dnn.readNet(model='./weights/yolov3-tiny.weights',
                      config='./config/yolov3-tiny.cfg')
# 由yolo-v3的结构可知，最终有三个尺度的输出
layerName = net.getLayerNames()
# 存储输出的三个尺度名称，用于后面进行前向推断的
ThreeOutput_layers_name = []
for i in net.getUnconnectedOutLayers():
    ThreeOutput_layers_name.append(layerName[i - 1])

# 因为yolo-v3中检测包含80个类别，所以首先获取类别
with open('./data/coco.names', 'r') as fp:
    classes = fp.read().splitlines()

# 指定过滤的置信度阈值：confidence
Confidence_thresh = 0.2
# 指定非极大值抑制的值：对候选框进行筛选
Nms_thresh = 0.35


# 检测的过程已经图形的绘制
def Forward_Predict(frame):
    # 参数情况：图像 ，归一化，缩放的大小，是否对RGB减去一个常数，R和B交换（因为R和B是反着的，所以需要交换），是否裁剪
    blob = cv2.dnn.blobFromImage(frame, 1 / 255, (64, 64), (0, 0, 0), swapRB=True, crop=False)
    # 获取图像的高宽
    height, width, channel = frame.shape
    # 设置网络输入
    net.setInput(blob)
    # 进行前向推断:采用的最后三个尺度输出层作为前向推断
    predict = net.forward(ThreeOutput_layers_name)
    # 存放预测框的坐标
    boxes = []
    # 存在预测物体的置信度
    confid_object = []
    # 存放预测的类别
    class_prob = []
    # 存放预测物体的id
    class_id = []
    # 存放预测类别的名称
    class_names = []
    # 根据输出的是三个尺度，所以分别遍历三个尺度
    for scale in predict:
        for box in scale:
            # 获取坐标值和高宽
            # 首先获取矩形中心坐标值（这里需要映射回原图）
            center_x = int(box[0] * width)
            center_y = int(box[1] * height)
            # 计算框的高宽
            w = int(box[2] * width)
            h = int(box[3] * height)
            # 获取矩形框的左上角坐标
            left_x = int(center_x - w / 2)
            left_y = int(center_y - h / 2)
            boxes.append([left_x, left_y, w, h])

            # 获取检测物体的置信度
            confid_object.append(float(box[4]))
            # 获取概率最大值
            # 首先获取最高值概率的下标
            index = np.argmax(box[5:])
            class_id.append(index)
            class_names.append(classes[index])
            class_prob.append(box[index])
    confidences = np.array(class_prob) * np.array(confid_object)
    # 计算非极大值抑制
    all_index = cv2.dnn.NMSBoxes(boxes, confidences, Confidence_thresh, Nms_thresh)

    try:
        for i in all_index.flatten():
            x, y, w, h = boxes[i]
            # 四舍五入，保留2位小数
            confidence = str(round(confidences[i], 2))
            # 绘制矩形框
            cv2.rectangle(img=frame, pt1=(x, y), pt2=(x + w, y + h),
                          color=(0, 255, 0), thickness=2)
            text = class_names[i] + ' ' + confidence
            cv2.putText(img=frame, text=text, org=(x, y - 10),
                        fontFace=cv2.FONT_HERSHEY_SIMPLEX,
                        fontScale=1.0, color=(0, 0, 255), thickness=2)
    except IOError:
        pass
    return frame


# 实时的检测
def detect_time():
    cap = cv2.VideoCapture(0)
    while cap.isOpened():
        OK, frame = cap.read()
        if not OK:
            break
        # 将图片进行一下翻转，因为Opencv读取的图片和我们正常是反着的
        frame = cv2.flip(src=frame, flipCode=2)
        frame = cv2.resize(src=frame, dsize=(64, 64))

        try:
            t1 = time.time()
            dst = Forward_Predict(frame)
            cv2.namedWindow("detect", cv2.WINDOW_NORMAL)
            cv2.imshow('detect', dst)
            t2 = time.time()
            print(f'Done. ({(1E3 * (t2 - t1)):.1f}ms) Inference')
        except :
            cv2.imshow('detect', frame)
        key = cv2.waitKey(1)
        if key == 27:
            break
    cap.release()


# 单张图片的检测
def signa_Picture(image_path='images/smile.jpg'):
    img = cv2.imread(image_path)
    img = cv2.resize(src=img, dsize=(416, 416))
    dst = Forward_Predict(img)
    cv2.imshow('detect', dst)
    key = cv2.waitKey(0)
    if key == 27:
        exit()


cv2.destroyAllWindows()

if __name__ == '__main__':
    print('Pycharm')
    # signa_Picture()
    detect_time()