非局部均值滤波算法的python实现

图像去噪算法

最新推荐文章于 2024-04-14 22:46:26 发布

转载最新推荐文章于 2024-04-14 22:46:26 发布 · 1.2k 阅读

图像专栏收录该内容

2 篇文章

订阅专栏

本文介绍了一种基于非局部均值滤波（NL-means）的图像去噪算法，并与中值滤波及OpenCV自带的快速非局部均值去噪算法进行了比较。通过计算峰值信噪比（PSNR），验证了不同去噪方法的效果。

部署运行你感兴趣的模型镜像

#coding:utf8
import cv2
import numpy as np
def psnr(A, B):
    return 10*np.log(255*255.0/(((A.astype(np.float)-B)**2).mean()))/np.log(10)

def double2uint8(I, ratio=1.0):
    return np.clip(np.round(I*ratio), 0, 255).astype(np.uint8)

def make_kernel(f):
    kernel = np.zeros((2*f+1, 2*f+1))
    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 NLmeansfilter(I, h_=10, templateWindowSize=5,  searchWindowSize=11):
    f = templateWindowSize/2
    t = searchWindowSize/2
    height, width = I.shape[:2]
    padLength = t+f
    I2 = np.pad(I, padLength, 'symmetric')
    kernel = make_kernel(f)
    h = (h_**2)
    I_ = I2[padLength-f:padLength+f+height, padLength-f:padLength+f+width]

    average = np.zeros(I.shape)
    sweight = np.zeros(I.shape)
    wmax =  np.zeros(I.shape)
    for i in range(-t, t+1):
        for j in range(-t, t+1):
            if i==0 and j==0:
                continue
            I2_ = I2[padLength+i-f:padLength+i+f+height, padLength+j-f:padLength+j+f+width]
            w = np.exp(-cv2.filter2D((I2_ - I_)**2, -1, kernel)/h)[f:f+height, f:f+width]
            sweight += w
            wmax = np.maximum(wmax, w)
            average += (w*I2_[f:f+height, f:f+width])
    return (average+wmax*I)/(sweight+wmax)

if __name__ == '__main__':
    I = cv2.imread('lena.jpg', 0)

    sigma = 20.0
    I1 = double2uint8(I + np.random.randn(*I.shape) *sigma)
    print u'噪声图像PSNR',psnr(I, I1)
    R1  = cv2.medianBlur(I1, 5)
    print u'中值滤波PSNR',psnr(I, R1)
    R2 = cv2.fastNlMeansDenoising(I1, None, sigma, 5, 11)
    print u'opencv的NLM算法',psnr(I, R2)
    R3 = double2uint8(NLmeansfilter(I1.astype(np.float), sigma, 5, 11))
    print u'NLM PSNR',psnr(I, R3)