如何使用TensorFlow Object Detection API提取bounding box内的图像？

你的代码已经实现了目标检测和可视化，但缺少提取检测框区域图像的逻辑。以下是修改后的完整代码，加入了裁剪并保存检测框内图像的功能：

while True:
    # 修正：cap是VideoCapture对象，应该用read()获取帧，而非imread()
    ret, frame = cap.read()
    if not ret:
        break

    # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
    expanded_frame = np.expand_dims(frame, axis=0)
    (boxes, scores, classes, num_c) = sess.run(
        [detection_boxes, detection_scores, detection_classes, num_detections],
        feed_dict={image_tensor: expanded_frame})

    # 获取图像尺寸，用于转换归一化坐标
    h, w, _ = frame.shape
    # 去除冗余维度
    boxes = np.squeeze(boxes)
    scores = np.squeeze(scores)
    classes = np.squeeze(classes).astype(np.int32)

    # 遍历检测结果，提取符合分数阈值的框内图像
    for idx in range(len(scores)):
        if scores[idx] >= 0.40:
            # 转换归一化坐标为实际像素坐标（注意boxes格式是[ymin, xmin, ymax, xmax]）
            ymin, xmin, ymax, xmax = boxes[idx]
            x1 = int(xmin * w)
            y1 = int(ymin * h)
            x2 = int(xmax * w)
            y2 = int(ymax * h)
            # 裁剪检测框区域
            cropped_image = frame[y1:y2, x1:x2]
            # 保存裁剪后的图像，用索引区分避免覆盖
            save_path = fr'C:\Users\user\Desktop\New folder\Face Detected\cropped_{idx}.png'
            cv2.imwrite(save_path, cropped_image)

    # Visualization of the detection
    vis_util.visualize_boxes_and_labels_on_image_array(
        frame,
        np.squeeze(boxes),
        classes,
        scores,
        category_index,
        use_normalized_coordinates=True,
        line_thickness=1,
        min_score_thresh=0.40)

    cv2.imshow('Detection', frame)
    filename = r'C:\Users\user\Desktop\New folder\Face Detected\Detected.png'
    cv2.imwrite(filename, frame)  
    fps.update()

    if cv2.waitKey(1) == ord('q'):
        fps.stop()
        break

print("Fps: {:.2f}".format(fps.fps()))
cap.release()  # 修正：用release()而非stop()，VideoCapture对象没有stop()方法
cv2.destroyAllWindows()

关键说明： #

• 坐标转换：TensorFlow Detection API返回的boxes是归一化坐标（范围0-1），需要乘以图像的实际宽高转换为像素坐标，注意坐标顺序是[ymin, xmin, ymax, xmax]
• 筛选有效框：只处理分数大于min_score_thresh（这里是0.40）的检测结果，避免无效框
• 避免文件覆盖：保存裁剪图像时加入索引idx，防止每次循环的图像互相覆盖
• 代码修正：原代码中frame = cv2.imread(cap)是错误的，cap作为视频捕获对象应该用cap.read()获取帧；cap.stop()改为cap.release()，符合OpenCV的API规范

内容的提问来源于stack exchange，提问作者Pooja