经典非局部均值滤波(NLM)算法python实现(2)

经典非局部均值滤波(NLM)算法python实现（三通道图像版本）

单通道图像版本已发布：

https://blog.csdn.net/yy0722a/article/details/113924087

以下用python(3.x)写的非局部均值滤波代码：

import cv2 
import numpy as np

# f为相似窗口的半径, t为搜索窗口的半径, h为高斯函数平滑参数(一般取为相似窗口的大小)
def make_kernel(f):
    kernel = np.zeros((2*f+1, 2*f+1), np.float32)
    for d in range(1, f+1):
        kernel[f-d:f+d+1, f-d:f+d+1] += (1.0/((2*d+1)**2))

    return kernel/kernel.sum()
    
def NLmeans_filter2(src, f, t, h):
    H, W, C = src.shape
    out = np.zeros((H, W, C), np.uint8)
    pad_length = f+t
    one, two, three = cv2.split(src)
    src_padding_one = np.pad(one, (pad_length, pad_length),mode='symmetric').astype(np.float32)
    src_padding_two = np.pad(two, (pad_length, pad_length),mode='symmetric').astype(np.float32)
    src_padding_three = np.pad(three,(pad_length, pad_length),mode='symmetric').astype(np.float32)
    src_padding = cv2.merge([src_padding_one, src_padding_two, src_padding_three])
    kernel = make_kernel(f)
    h2 = h*h

    for i in range(H):
        for j in range(W):
            for ch in range(C):
                i1 = i + f + t # 将未填充图像上的点坐标映射到填充后图像上的对应坐标
                j1 = j + f + t
                W1 = src_padding[i1-f:i1+f+1, j1-f:j1+f+1, ch] # 领域窗口W1
                w_max = 0
                aver = 0
                weight_sum = 0
            # 在搜索窗口进行搜索相似图像快
                for r in range(i1-t, i1+t+1):
                    for c in range(j1-t, j1+t+1):
                        if (r==i1) and (c==j1): # 自身图像块领域先不处理
                            continue
                        else:
                            W2 = src_padding[r-f:r+f+1, c-f:c+f+1, ch] # 搜索区域内的相似窗口W2
                            Dist2 = (kernel*(W1-W2)*(W1-W2)).sum()
                            w = np.exp(-Dist2/h2)
                            if w > w_max:
                                w_max = w
                            weight_sum += w
                            aver += w*src_padding[r, c, ch]
                aver += w_max*src_padding[i1, j1, ch] # 自身领域取最大的权重
                weight_sum += w_max
                out[i, j, ch] = aver/weight_sum

    return out
    
img = cv2.imread("lena256gs10.bmp")
out = NLmeans_filter2(img, 2, 5, 10)
cv2.imwrite("result2.bmp", out)

滤波结果：

原图像(加了标准差为10的高斯噪声的lena图像(256x256x3))：

滤波后的图像：

由图可见，高斯噪声基本被滤除， 图像纹理细节也得到了较好的保留。