带掩码的 Python OpenCV matchTemplate - 在所有位置找到匹配项

Question

问题:我从 matchTemplate 收到的结果表明我在每个位置都有匹配值 1.0。

预期结果:我预计 results 中的一个位置比其他位置有更高的分数。

代码:

def template_match(filename=base_name,
                   img_folder=trn_imgs_path,
                   templates=['wet_install.png',
                              'wet_install_cleaned.png',
                              'wet_install_tag.png',
                              'wet_install_tag_cleaned.png'],
                   template_path=template_path,
                   threshold=0.8,
                   save_dir=save_dir):
    ''' 
    Perform template matching on an input image using a few templates.
    It draws bounding boxes on a copy of the original image. 

    Args:
        filename (str): name of the file with the .svg extension
        img_folder (str): path to folder containing the images
        templates (list): list of template filenames to match against
        template_path (str): path to folder containing the templates
        threshold (float): the threshold for a match from template matching
        save_dir (str): path to folder to save results
    '''
    print('Working on file: {}.png'.format(filename))

    # load the original BGR image
    img_rgb  = cv2.imread(img_folder + filename + '.png')[5143:5296, 15169:15368] # TODO(mtu): Don't keep these indices here!
    img_gray = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
    img_gray = cv2.adaptiveThreshold(img_gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 115, 1)

    # loop over each template
    colors = [(0,0,255), (0,255,0), (255,255,0), (255,0,255)]
    for itemp in range(len(templates)):
        template_name = templates[itemp]
        print('Using Template: {}'.format(template_name))

        # load the template as grayscale and get its width and height
        template = cv2.imread(template_path + '{}'.format(template_name), 0)
        height, width = template.shape[:2]

        template  = cv2.adaptiveThreshold(template, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 115, 1)
        temp_mask = cv2.adaptiveThreshold(template, 1, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 115, 1)

        # do template matching using grayscale image and find points above theshold
        results = cv2.matchTemplate(image=img_gray, templ=template, method=cv2.TM_CCORR_NORMED, mask=temp_mask)
        loc = np.where(results >= threshold)

        # draw rectangles on points above threshold on RGB image
        for pt in zip(*loc[::-1]):
            cv2.rectangle(img_rgb, pt, (pt[0] + width, pt[1] + height), colors[itemp%len(colors)], 5)

    # save the file with bounding boxes drawn on
    filename = save_dir + '{}_found.png'.format(filename) 
    print('Saving bounding boxes to: {}'.format(filename))
    cv2.imwrite(filename, img_rgb)

评论:

• Debug output显示 img_gray、template 和 temp_mask 的外观
• img_gray 只是模板，顶部有 10 行额外的白色填充像素
• template 和temp_mask 是相同的形状和类型
• Saved output image

Answer 1

反转 img_gray 和 template 修复了错误。

我使用的比较指标是 cv2.TM_CCORR_NORMED。这通过获取 img_gray 和 template 的点积来实现，其中二进制 numpy 数组 temp_mask 的值为 1。

在我的示例图像中，我想将 template 中的黑色像素与 img_gray 中的黑色像素进行匹配，但是黑色的像素值为 0。因此，我想要检测的位置的点积很低。

通过反转 img_gray 和 template 我将 template 中的白色像素与 img_gray 中的白色像素进行匹配.由于白色的像素值为 255，因此白色与白色、模板与图像的点积在我想要检测的位置变得很高。