MATLAB与Python图像显著图差异：暗斑问题的原因及解决方案

问题背景 #

针对一张450x620的图像，通过3x3中值滤波与35x35均值滤波结果相减取绝对值的方式计算显著图。MATLAB实现可得到正常输出，但使用PyTorch+OpenCV在Colab中实现时，部分区域的均值滤波与中值滤波像素值完全相同，导致生成图像出现暗斑。

可能原因 #

• 数据类型与精度差异：MATLAB中imread读取灰度图后转double是0-255的浮点值，滤波全程基于浮点运算，精度保留完整；而OpenCV的cv.imread默认读取为uint8整数型，后续滤波直接做整数运算，会丢失小数精度，平缓区域的滤波结果易完全重合。
• 边界填充策略不一致：MATLAB的imfilter使用'replicate'（复制边界像素）填充方式，而OpenCV的cv.blur、cv.medianBlur默认采用零填充。尤其35x35大核均值滤波时，边界区域的均值会被零拉低，与中值滤波结果更易出现完全相等的情况。
• 滤波计算逻辑差异：OpenCV均值滤波直接对整数像素求和取整，中值滤波也仅返回整数结果；MATLAB则全程保持浮点运算，即使是平缓区域，滤波结果也会保留微小的小数差异，不会出现完全相等的情况。

解决方案 #

1. 统一浮点数据类型：读取图像后立即转为浮点型，确保滤波运算在高精度下进行：

   import numpy as np
   import cv2
   
   im1 = cv2.imread('/content/IR12.png', cv2.IMREAD_GRAYSCALE).astype(np.float64)
   im2 = cv2.imread('/content/VIS12.png', cv2.IMREAD_GRAYSCALE).astype(np.float64)
   

2. 对齐边界填充策略：用cv2.filter2D替代cv2.blur，手动设置borderType=cv2.BORDER_REPLICATE，和MATLAB的边界处理逻辑保持一致；中值滤波可使用scipy.ndimage.median_filter自定义边界填充：

   from scipy.ndimage import median_filter
   
   # 3x3均值滤波（复制边界）
   kernel_3 = np.ones((3,3), np.float64) / 9
   base1 = cv2.filter2D(im1, -1, kernel_3, borderType=cv2.BORDER_REPLICATE)
   # 35x35均值滤波（复制边界）
   kernel_35 = np.ones((35,35), np.float64) / (35*35)
   meanIR = cv2.filter2D(im1, -1, kernel_35, borderType=cv2.BORDER_REPLICATE)
   # 3x3中值滤波（复制边界）
   medIR = median_filter(im1, size=3, mode='nearest')
   

3. 避免权重计算异常：计算权重时给分母添加极小值epsilon，防止salIR + salVIS为0导致的暗斑问题：

   eps = 1e-8
   weightIR = salIR / (salIR + salVIS + eps)
   weightVIS = salVIS / (salIR + salVIS + eps)
   

4. 后处理修正：生成融合图像后，将数值钳位到0-255范围并转为uint8，确保可视化效果正常：

   fused_np = np.clip(fused.numpy(), 0, 255).astype(np.uint8)
   cv2_imshow(fused_np)
   

代码对比与输出 #

修正前的Google Colab PyTorch代码 #

# Load the IR and visible images
im1 = cv.imread('/content/IR12.png')
im1 = cv.cvtColor(im1, cv.COLOR_RGB2GRAY)
im2 = cv.imread('/content/VIS12.png')
im2 = cv.cvtColor(im2, cv.COLOR_RGB2GRAY)
IR = torch.from_numpy(im1).double()
VIS = torch.from_numpy(im2).double()

# Apply mean filtering
base1 = cv.blur(im1,(3,3))
base2 = cv.blur(im2,(3,3))
base1 = torch.from_numpy(base1).double()
base2 = torch.from_numpy(base2).double()

# Compute detail maps
detail1 = IR - base1
detail2 = VIS - base2
#cv2_imshow(detail1)

# Apply median filtering
medIR = cv.medianBlur(im1, 3)
medIR = torch.from_numpy(medIR).double()
medVIS = cv.medianBlur(im2, 3)
medVIS = torch.from_numpy(medVIS).double()

# Apply mean filtering using a larger kernel
meanIR = cv.blur(im1,(35,35))
meanIR = torch.from_numpy(meanIR).double()
meanVIS = cv.blur(im2,(35,35))
meanVIS = torch.from_numpy(meanVIS).double()


# Compute saliency maps
salIR = abs(meanIR - medIR)
cv2_imshow(salIR.numpy())
salVIS = abs(meanVIS - medVIS)
#salIR = torch.from_numpy(salIR).double()
#salVIS = torch.from_numpy(salVIS).double()


# Compute weight maps
weightIR = salIR / (salIR + salVIS)
weightVIS = salVIS / (salIR + salVIS)

# Fuse the base and detail maps
basef = 0.5 * (base1 + base2)
detailf = (weightIR * detail1) + (weightVIS * detail2)

# Compute the fused image
fused = torch.add(basef,detailf)

# Convert the fused image to a NumPy array
fused_np = fused.numpy()
cv2_imshow(fused_np) 

修正前Python输出

MATLAB代码 #

Im2 = imread('D:\ankan\IV_images\IV_images\VIS12.png');
Im1 = imread('D:\ankan\IV_images\IV_images\IR12.png');
IR = double(Im1);
VIS = double(Im2);
[m,n] = size(Im2);

kernel1 = 3;
kernel2 = 35;

meanFilter1 = ones(kernel1)/(kernel1^2);
meanFilter2 = ones(kernel2)/(kernel2^2);

base1 = imfilter (IR,meanFilter1,'conv','replicate');
base2 = imfilter (VIS,meanFilter1,'conv','replicate');

detail1 = IR-base1;
detail2 = VIS-base2;

meanIR = imfilter (IR,meanFilter2,'conv','replicate');
meanVIS = imfilter (VIS,meanFilter2,'conv','replicate');

medIR = medfilt2(IR,[3,3]);
medVIS = medfilt2(VIS,[3,3]);

salIR = abs(meanIR - medIR);
salVIS = abs(meanVIS - medVIS);

weightIR = salIR./(salIR+salVIS);
weightVIS = salVIS./(salIR+salVIS);

basef = 0.5*(base1+base2);
detailf = (weightIR.*detail1)  +  (weightVIS.*detail2);

fused = basef+detailf;
imshow(uint8(fused))

MATLAB输出

内容的提问来源于stack exchange，提问作者Ankan Banerjee