参考文章：https://blog.csdn.net/u010839382/article/details/48241929 https://blog.csdn.net/u010839382/article/details/48229579
原理方面就不详述了，参考文章里都有。参考的两篇文章都是用matlab实现的，代码读起来比较简单。博主用python3.7实现了改进加速后的算法，发现耗时与matlab天差地别，慢的不行，说明matlab矩阵运算很厉害，自己对于python提速的理解也不够到位。不过通过将代码从matlab改写为python还是学到了不少东西，分享出来希望给大家提供点思路。
代码就是这三个函数，里面的arraycompare应该是能改进的，没找到python有啥能够直接将两个矩阵中相对应位置元素最大值返回的函数，matlab里面直接max(矩阵A，矩阵B)就完事了T _ T。
def arraycompare(array1, array2, height, width): resultarray = np.zeros((height, width)) for row in range(0, height): for col in range(0, width): resultarray[row, col] = max(array1[row, col], array2[row, col]) return resultarray def integralImgSqDiff2(paddedimg_val, Ds_val, t1_val, t2_val): lengthrow = len(paddedimg_val[:, 0]) lengthcol = len(paddedimg_val[0, :]) Dist2 = (paddedimg_val[Ds_val:-Ds_val, Ds_val:-Ds_val] - paddedimg_val[Ds_val + t1_val:lengthrow - Ds_val + t1_val, Ds_val + t2_val:lengthcol - Ds_val + t2_val]) ** 2 Sd_val = Dist2.cumsum(0) Sd_val = Sd_val.cumsum(1) return Sd_val def nl_meansfilter(imagearray, h_=10, ds0=2, ds1=5): height, width = imagearray[:, :, 0].shape[:2] length0 = height + 2 * ds1 length1 = width + 2 * ds1 h = (h_ ** 2) d = (2 * ds0 + 1) ** 2 imagearray_NL = np.zeros(imagearray.shape).astype('uint8') for i in range(0, 3): paddedimg = np.pad(imagearray[:, :, i], ds0 + ds1 + 1, 'symmetric') paddedimg = paddedimg.astype('float64') paddedv = np.pad(imagearray[:, :, i], ds1, 'symmetric') paddedv = paddedv.astype('float64') average = np.zeros((height, width)) sweight = np.zeros((height, width)) wmax = np.zeros((height, width)) for t1 in range(-ds1, ds1 + 1): for t2 in range(-ds1, ds1 + 1): if t1 == 0 and t2 == 0: continue Sd = integralImgSqDiff2(paddedimg, ds1, t1, t2) SqDist2 = Sd[2 * ds0 + 1:-1, 2 * ds0 + 1:-1] + Sd[0:-2 * ds0 - 2, 0:-2 * ds0 - 2] - \ Sd[2 * ds0 + 1:-1, 0:-2 * ds0 - 2] - Sd[0:-2 * ds0 - 2, 2 * ds0 + 1:-1] SqDist2 /= d * h w = np.exp(-SqDist2) v = paddedv[ds1 + t1:length0 - ds1 + t1, ds1 + t2:length1 - ds1 + t2] average += w * v wmax = arraycompare(wmax, w, height, width) sweight += w average += wmax * imagearray[:, :, i] average /= wmax + sweight average_uint8 = average.astype('uint8') imagearray_NL[:, :, i] = average_uint8 return imagearray_NL