下載并將YOLOv5模型轉(zhuǎn)換為onnx格式

YOLOv5 v6.1版本合并了輸出，更加容易解析輸出結(jié)果

參考https://docs.ultralytics.com/quick-start/艺糜，完成YOLOv5 開發(fā)環(huán)境搭建

使用命令將YOLOv5 Pytorch模型轉(zhuǎn)換為onnx模型：

python export.py --weights yolov5s.pt --include onnx

1647659625(1).png

網(wǎng)絡(luò)原因术幔，從Github上下載YOLOv5權(quán)重失敗

使用訓(xùn)練解決下載YOLOv5權(quán)重問題

轉(zhuǎn)換onnx模型成功

使用Netron查看yolov5s.onnx v6.1模型

yolov5s.onnx v6.1模型

可以看到：

• 模型輸入節(jié)點，name：images; shape NCHW=[1,3,640,640]; type:float32
• 模型輸出節(jié)點屠尊，name: output; shape [1,25200,85], type:float32

output整合了之前三層(v6.1之前版本是三個輸出節(jié)點)的原始輸出，每一行85個數(shù)值耕拷，前面5個數(shù)值分別是：

cx, cy, w, h, score 后面80個MSCOCO的分類得分

使用OpenCV DNN API完成推理

完整GitHub Repo 地址：https://gitee.com/ppov-nuc/yolov5_infer

import cv2
import numpy as np
import time
import yaml

# 載入COCO Label
with open('./coco.yaml','r', encoding='utf-8') as f:
    result = yaml.load(f.read(),Loader=yaml.FullLoader)
class_list = result['names']

# YOLOv5s輸入尺寸
INPUT_WIDTH = 640
INPUT_HEIGHT = 640

# 目標檢測函數(shù)讼昆，返回檢測結(jié)果
def detect(image, net):
    blob = cv2.dnn.blobFromImage(image, 1 / 255.0, (INPUT_WIDTH, INPUT_HEIGHT), swapRB=True, crop=False)
    net.setInput(blob)
    preds = net.forward()
    return preds

# YOLOv5的后處理函數(shù)，解析模型的輸出
def wrap_detection(input_image, output_data):
    class_ids = []
    confidences = []
    boxes = []
    #print(output_data.shape)
    rows = output_data.shape[0]

    image_width, image_height, _ = input_image.shape

    x_factor = image_width / INPUT_WIDTH
    y_factor = image_height / INPUT_HEIGHT

    for r in range(rows):
        row = output_data[r]
        confidence = row[4]
        if confidence >= 0.4:

            classes_scores = row[5:]
            _, _, _, max_indx = cv2.minMaxLoc(classes_scores)
            class_id = max_indx[1]
            if (classes_scores[class_id] > .25):
                confidences.append(confidence)

                class_ids.append(class_id)

                x, y, w, h = row[0].item(), row[1].item(), row[2].item(), row[3].item()
                left = int((x - 0.5 * w) * x_factor)
                top = int((y - 0.5 * h) * y_factor)
                width = int(w * x_factor)
                height = int(h * y_factor)
                box = np.array([left, top, width, height])
                boxes.append(box)

    indexes = cv2.dnn.NMSBoxes(boxes, confidences, 0.25, 0.45)

    result_class_ids = []
    result_confidences = []
    result_boxes = []

    for i in indexes:
        result_confidences.append(confidences[i])
        result_class_ids.append(class_ids[i])
        result_boxes.append(boxes[i])

    return result_class_ids, result_confidences, result_boxes

# 按照YOLOv5要求骚烧，先將圖像長:寬 = 1:1浸赫，多余部分填充黑邊
def format_yolov5(frame):
    row, col, _ = frame.shape
    _max = max(col, row)
    result = np.zeros((_max, _max, 3), np.uint8)
    result[0:row, 0:col] = frame
    return result


# 載入yolov5s.onnx模型
model_path = "./yolov5s.onnx"
# Read yolov5s.onnx model with OpenCV API
net = cv2.dnn.readNetFromONNX(model_path)

# 開啟Webcam，并設(shè)置為1280x720
cap = cv2.VideoCapture(0,cv2.CAP_DSHOW)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 1280)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)

# 調(diào)色板
colors = [(255, 255, 0), (0, 255, 0), (0, 255, 255), (255, 0, 0)]

# 開啟檢測循環(huán)
while True:
    start = time.time()
    _, frame = cap.read()
    if frame is None:
        print("End of stream")
        break
    # 將圖像按最大邊1:1放縮
    inputImage = format_yolov5(frame)
    # 執(zhí)行推理計算
    outs = detect(inputImage, net)
    # 拆解推理結(jié)果
    class_ids, confidences, boxes = wrap_detection(inputImage, outs[0])

    # 顯示檢測框bbox
    for (classid, confidence, box) in zip(class_ids, confidences, boxes):
        color = colors[int(classid) % len(colors)]
        cv2.rectangle(frame, box, color, 2)
        cv2.rectangle(frame, (box[0], box[1] - 20), (box[0] + box[2], box[1]), color, -1)
        cv2.putText(frame, class_list[classid], (box[0], box[1] - 10), cv2.FONT_HERSHEY_SIMPLEX, .5, (0, 0, 0))
    
    # 顯示推理速度FPS
    end = time.time()
    inf_end = end - start
    fps = 1 / inf_end
    fps_label = "FPS: %.2f" % fps
    cv2.putText(frame, fps_label, (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
    print(fps_label+ "; Detections: " + str(len(class_ids)))
    cv2.imshow("output", frame)

    if cv2.waitKey(1) > -1:
        print("finished by user")
        break

使用OpenVINO API完成推理

import cv2
import numpy as np
import time
import yaml
from openvino.inference_engine import IECore # the version of openvino <= 2021.4.2

# 載入COCO Label
with open('./coco.yaml','r', encoding='utf-8') as f:
    result = yaml.load(f.read(),Loader=yaml.FullLoader)
class_list = result['names']

# YOLOv5s輸入尺寸
INPUT_WIDTH = 640
INPUT_HEIGHT = 640

# 目標檢測函數(shù)赃绊，返回檢測結(jié)果
def detect(image, net):
    blob = cv2.dnn.blobFromImage(image, 1 / 255.0, (INPUT_WIDTH, INPUT_HEIGHT), swapRB=True, crop=False)
    result = net.infer({"images": blob})
    preds = result["output"]
    return preds

# YOLOv5的后處理函數(shù)既峡，解析模型的輸出
def wrap_detection(input_image, output_data):
    class_ids = []
    confidences = []
    boxes = []
    #print(output_data.shape)
    rows = output_data.shape[0]

    image_width, image_height, _ = input_image.shape

    x_factor = image_width / INPUT_WIDTH
    y_factor = image_height / INPUT_HEIGHT

    for r in range(rows):
        row = output_data[r]
        confidence = row[4]
        if confidence >= 0.4:

            classes_scores = row[5:]
            _, _, _, max_indx = cv2.minMaxLoc(classes_scores)
            class_id = max_indx[1]
            if (classes_scores[class_id] > .25):
                confidences.append(confidence)

                class_ids.append(class_id)

                x, y, w, h = row[0].item(), row[1].item(), row[2].item(), row[3].item()
                left = int((x - 0.5 * w) * x_factor)
                top = int((y - 0.5 * h) * y_factor)
                width = int(w * x_factor)
                height = int(h * y_factor)
                box = np.array([left, top, width, height])
                boxes.append(box)

    indexes = cv2.dnn.NMSBoxes(boxes, confidences, 0.25, 0.45)

    result_class_ids = []
    result_confidences = []
    result_boxes = []

    for i in indexes:
        result_confidences.append(confidences[i])
        result_class_ids.append(class_ids[i])
        result_boxes.append(boxes[i])

    return result_class_ids, result_confidences, result_boxes

# 按照YOLOv5要求，先將圖像長:寬 = 1:1碧查，多余部分填充黑邊
def format_yolov5(frame):
    row, col, _ = frame.shape
    _max = max(col, row)
    result = np.zeros((_max, _max, 3), np.uint8)
    result[0:row, 0:col] = frame
    return result


# 載入yolov5s onnx模型
model_path = "./yolov5s.onnx"
# Read yolov5s onnx model with OpenVINO API
ie = IECore()
exec_net = ie.load_network(network=model_path, device_name="CPU")

# 開啟Webcam运敢，并設(shè)置為1280x720
cap = cv2.VideoCapture(0,cv2.CAP_DSHOW)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 1280)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)

# 調(diào)色板
colors = [(255, 255, 0), (0, 255, 0), (0, 255, 255), (255, 0, 0)]

# 開啟檢測循環(huán)
while True:
    start = time.time()
    _, frame = cap.read()
    if frame is None:
        print("End of stream")
        break
    # 將圖像按最大邊1:1放縮
    inputImage = format_yolov5(frame)
    # 執(zhí)行推理計算
    outs = detect(inputImage, exec_net)
    # 拆解推理結(jié)果
    class_ids, confidences, boxes = wrap_detection(inputImage, outs[0])

    # 顯示檢測框bbox
    for (classid, confidence, box) in zip(class_ids, confidences, boxes):
        color = colors[int(classid) % len(colors)]
        cv2.rectangle(frame, box, color, 2)
        cv2.rectangle(frame, (box[0], box[1] - 20), (box[0] + box[2], box[1]), color, -1)
        cv2.putText(frame, class_list[classid], (box[0], box[1] - 10), cv2.FONT_HERSHEY_SIMPLEX, .5, (0, 0, 0))
    
    # 顯示推理速度FPS
    end = time.time()
    inf_end = end - start
    fps = 1 / inf_end
    fps_label = "FPS: %.2f" % fps
    cv2.putText(frame, fps_label, (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
    print(fps_label+ "; Detections: " + str(len(class_ids)))
    cv2.imshow("output", frame)

    if cv2.waitKey(1) > -1:
        print("finished by user")
        break