OpenCV4.4加载TensorFlow1.x版本 Mask R-CNN模型进行推理

opencv中dnn模块的接口可以加载caffe，TensorFlow，darknet和onnx等模型，虽然目前pytorch非常流行，但由于pytorch属于动态图，部署的时候
没有静态图方便，虽然可以转换为onnx模型，但个人使用经验来看仍然存在很多接口兼容性问题。使用cv2.dnn.readNetFromTensorFlow加载tf模型
算是目前比较好的方法。下面记录采用cv2.dnn.readNetFromTensorFlow接口加载tf1.X版本的预训练模型并实现推理（似乎还不支持tf2.x版本），下载地址为：
http://download.tensorflow.org/models/object_detection/mask_rcnn_inception_v2_coco_2018_01_28.tar.gz

1. 解压tf1.x版本mask rcnn预训练模型
   

2. 生成.pbtxt描述文件

   OpenCV DNN需要根据.pbtxt模型描述文件来解析Tensorflow的pb模型文件，实现网络模型的加载。OpenCV源码中提供了生成.pbtxt文件的python脚本文件，文件位于OpenCV安装路径\opencv\sources\samples\dnn文件夹下，根据模型类别选择相应的文件即可。
   

   执行下列命令：

    python tf_text_graph_mask_rcnn.py --input "D:\GoogleDownload\mask_rcnn_inception_v2_coco_2018_01_28\frozen_inference_graph.pb" --output "D:\GoogleDownload\mask_rcnn_inception_v2_coco_2018_01_28\mask_rcnn_inceptionv2_coco.pbtxt" --config "D:\GoogleDownload\mask_rcnn_inception_v2_coco_2018_01_28\pipeline.config"
   

3. 模型推理。笔者使用的是dnn.readNet()接口，本来想像目标检测那样使用cv2.dnn_SegmentationModel接口，但是报错了，还不知道问题所在，因此参考了官方例子mask_rcnn.py进行实现。

    # -*- coding: utf-8 -*-
    """
    Created on Fri Aug 14 22:21:53 2020
    
    @author: 周文青
    
    opencv dnn模块加载mask rcnn模型
    """
    
    import cv2
    import cv2 as cv    
    import matplotlib.pyplot as plt
    import numpy as np
    coco_names = r"F:\opencv\sources\samples\dnn\coco.names"
    
    args = {
    'model':r"D:\GoogleDownload\mask_rcnn_inception_v2_coco_2018_01_28\frozen_inference_graph.pb",
    'config':r"D:\GoogleDownload\mask_rcnn_inception_v2_coco_2018_01_28\mask_rcnn_inceptionv2_coco.pbtxt",
    'width':800,'height':800,'thr':0.5}
    
    img_file = r"C:\Users\admin\Pictures\car.jpg"
    
    with open(coco_names,'rt') as f:
        classes = f.read().rstrip('\n').split('\n')
    
    frame=cv2.imread(img_file)
    
    # Load a network
    net = cv.dnn.readNet(cv.samples.findFile(args['model']), cv.samples.findFile(args['config']))
    net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
    
    winName = 'Mask-RCNN in OpenCV'
    # cv.namedWindow(winName, cv.WINDOW_NORMAL)
    
    frameH = frame.shape[0]
    frameW = frame.shape[1]
    
    # Create a 4D blob from a frame.
    blob = cv.dnn.blobFromImage(frame, size=(args['width'], args['height']), swapRB=True, crop=False)
    
    # Run a model
    net.setInput(blob)
    
    boxes, masks = net.forward(['detection_out_final', 'detection_masks'])
    
    numClasses = masks.shape[1]
    numDetections = boxes.shape[2]
    
    # Load colors
    colors = None
    if colors:
        with open(colors, 'rt') as f:
            colors = [np.array(color.split(' '), np.uint8) for color in f.read().rstrip('\n').split('\n')]
            
    # Draw segmentation
    if not colors:
        # Generate colors
        colors = [np.array([0, 0, 0], np.uint8)]
        for i in range(1, numClasses + 1):
            colors.append((colors[i - 1] + np.random.randint(0, 256, [3], np.uint8)) / 2)
        del colors[0]
        
    def drawBox(frame, classId, conf, left, top, right, bottom):
        # Draw a bounding box.
        cv.rectangle(frame, (left, top), (right, bottom), (0, 255, 0))
    
        label = '%.2f' % conf
    
        # Print a label of class.
        if classes:
            assert(classId < len(classes))
            label = '%s: %s' % (classes[classId], label)
    
        labelSize, baseLine = cv.getTextSize(label, cv.FONT_HERSHEY_SIMPLEX, 0.5, 1)
        top = max(top, labelSize[1])
        cv.rectangle(frame, (left, top - labelSize[1]), (left + labelSize[0], top + baseLine), (255, 255, 255), cv.FILLED)
        cv.putText(frame, label, (left, top), cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0))
        
    boxesToDraw = []
    for i in range(numDetections):
        box = boxes[0, 0, i]
        mask = masks[i]
        score = box[2]
        if score > args['thr']:
            classId = int(box[1])
            left = int(frameW * box[3])
            top = int(frameH * box[4])
            right = int(frameW * box[5])
            bottom = int(frameH * box[6])
    
            left = max(0, min(left, frameW - 1))
            top = max(0, min(top, frameH - 1))
            right = max(0, min(right, frameW - 1))
            bottom = max(0, min(bottom, frameH - 1))
    
            boxesToDraw.append([frame, classId, score, left, top, right, bottom])
    
            classMask = mask[classId]
            classMask = cv.resize(classMask, (right - left + 1, bottom - top + 1))
            mask = (classMask > 0.5)
    
            roi = frame[top:bottom+1, left:right+1][mask]
            frame[top:bottom+1, left:right+1][mask] = (0.7 * colors[classId] + 0.3 * roi).astype(np.uint8)
    
    for box in boxesToDraw:
        drawBox(*box)
    
    # Put efficiency information.
    t, _ = net.getPerfProfile()
    label = 'Inference time: %.2f ms' % (t * 1000.0 / cv.getTickFrequency())
    cv.putText(frame, label, (0, 15), cv.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 0))
    
    plt.imshow(frame[:,:,::-1])
    plt.xticks([])
    plt.yticks([])
    plt.box(False)
    plt.tight_layout(pad=0.0)
    plt.show()
   

检测结果为：

通过调节阈值也可以得到更多的检测结果，只不过可行度有所下降。

另外这幅图在YOLOV3和YOLOV4模型下的检测结果也可以查看笔者的另一篇文章：
https://editor.csdn.net/md/?articleId=108015232

参考链接：
[1] https://blog.csdn.net/stjuliet/article/details/97294104