OpenCV Python与C++版imreconstruct实现结果不一致排查

我用Python和C分别实现了MATLAB的imreconstruct函数，测试发现Python版本输出与MATLAB一致，但C版本结果不符。测试案例中，Python输出的第一行第四、五列和第四、五行第一列保持mask的原始值3，而C++版本中这些位置的值被错误修改为2.7555513、2.3122778等。

Python实现代码 #

import numpy as np
import cv2 as cv

def imReconstruct(marker: np.array, mask: np.array) -> np.array:
    """
    Naive implementation of MatLAB's imReconstruct function
    works when `mask` consists of mostly background (global minimum)
    will be slow otherwise
    """
    kernel = cv.getStructuringElement(cv.MORPH_RECT, (3, 3))

    # calculate the extreme values of the mask image
    min_val, max_val, _, _ = cv.minMaxLoc(mask)

    # clip the marker by global extrema of mask
    _, marker = cv.threshold(marker, min_val, max_val, cv.THRESH_TRUNC | cv.THRESH_BINARY_INV)

    while True:
        expanded = cv.dilate(marker, kernel)
        expanded = np.minimum(expanded, mask)

        # return `expanded` when the difference is small
        if np.max(np.abs(expanded - marker)) < 1e-5:
            return expanded

        # set expanded to marker and repeat
        marker = expanded


D = np.array(
      [[4.2426405, 3.6055512, 3.1622777, 3.       , 3.       ],
       [3.6055512, 2.828427 , 2.236068 , 2.       , 2.       ],
       [3.1622777, 2.236068 , 1.4142135, 1.       , 1.       ],
       [3.       , 2.       , 1.       , 0.       , 0.       ],
       [3.       , 2.       , 1.       , 0.       , 0.       ]], dtype=np.float32)

# 输出结果
print(imReconstruct(D-.85,D))
# array([[3.3926406, 3.3926406, 3.1622777, 3.       , 3.       ],
#        [3.3926406, 2.828427 , 2.236068 , 2.       , 2.       ],
#        [3.1622777, 2.236068 , 1.4142135, 1.       , 1.       ],
#        [3.       , 2.       , 1.       , 0.       , 0.       ],
#        [3.       , 2.       , 1.       , 0.       , 0.       ]], dtype=float32)

C++实现代码 #

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

Mat imReconstruct(Mat marker, Mat mask){
    /**************
     * naive implementation of MatLAB's imReconstruct
     * works when `mask` consist of mostly background (global minimum)
     * will be slow otherwise
     *************/
    Mat kernel = getStructuringElement(MORPH_RECT, Size(3,3));

    // calculate the min and max values from mask
    double minMask, maxMask;
    minMaxLoc(mask, &minMask, &maxMask);

    // clip the marker by global extrema of mask
    threshold(marker, marker, minMask, maxMask, THRESH_TRUNC|THRESH_BINARY_INV);

    Mat expanded;

    // keep filling the holes with `dilate`
    // until there are no more changes
    while (1){
        dilate(marker, expanded, kernel);
        expanded = min(expanded, mask);

        // compute the max difference
        minMaxLoc(expanded-marker, &minMask, &maxMask);

        // return image when changes are small
        if (maxMask<1e-5) return expanded;

        // set expanded as marker and continue looping
        marker = expanded;
    }
}

int main() {
    // test case
    cv::Mat D = (cv::Mat_<float> (5,5) <<
           4.2426405, 3.6055512, 3.1622777, 3.       , 3.       ,
           3.6055512, 2.828427 , 2.236068 , 2.       , 2.       ,
           3.1622777, 2.236068 , 1.4142135, 1.       , 1.       ,
           3.       , 2.       , 1.       , 0.       , 0.       ,
           3.       , 2.       , 1.       , 0.       , 0.       
    );

    std::cout << imReconstruct(D - .85, D) << std::endl;
    // 输出结果
    // [3.3926406, 3.3926406, 3.1622777, 2.7555513, 2.236068;
    //  3.3926406, 2.8284271, 2.236068, 2, 2;
    //  3.1622777, 2.236068, 1.4142135, 1, 1;
    //  2.7555513, 2, 1, 0, 0;
    //  2.236068, 2, 1, 0, 0]
    return 0;
}

核心差异原因 #

1. 阈值操作的非法组合与行为差异 #

OpenCV的threshold函数的阈值类型（如THRESH_TRUNC、THRESH_BINARY_INV）是互斥的，不能同时指定多个。你试图用THRESH_TRUNC | THRESH_BINARY_INV实现marker的裁剪（限制在mask的min和max之间），但这个组合属于未定义行为：

• Python版本的OpenCV可能忽略了无效的flags组合，仅执行了THRESH_TRUNC，或巧合得到了类似裁剪的效果；
• C++版本的OpenCV对该组合的处理逻辑不同，导致初始marker的数值被错误修改，后续循环的膨胀和取min操作自然偏离预期。

正确的裁剪方式：

• Python中替换为marker = np.clip(marker, min_val, max_val)；
• C++中替换为：

  marker.setTo(minMask, marker < minMask);
  marker.setTo(maxMask, marker > maxMask);
  

2. 循环终止条件的逻辑不一致 #

• Python中计算的是np.max(np.abs(expanded - marker))，即差值绝对值的最大值，确保所有像素的变化都小于阈值；
• C++中通过minMaxLoc(expanded - marker, &minMask, &maxMask)仅获取了差值的最大值（正差值），忽略了负差值的绝对值可能超过1e-5的情况，导致循环提前终止，未完成完整的重建过程。

修正方式：C++中先对差值取绝对值，再计算最大值：

Mat diff;
absdiff(expanded, marker, diff);
minMaxLoc(diff, &minMask, &maxMask);
if (maxMask < 1e-5) return expanded;

内容的提问来源于stack exchange，提问作者Quang Hoang