Python图像高光调整

看了这个文章，里面有专门的c++的实现，我这边简单的使用python进行了实现，实现了两个版本，一个是python遍历像素，一个是使用numpy加速，代码如下：

import time
import numpy as np
import cv2def lighting(img, light):assert -100 <= light <= 100max_v = 4bright = (light/100.0)/max_vmid = 1.0+max_v*brightprint('bright: ', bright, 'mid: ', mid)gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY).astype(np.float32)/255.0thresh = gray*grayt = np.mean(thresh)mask = np.where(thresh > t, 255, 0).astype(np.float32)brightrate = np.zeros_like(mask).astype(np.float32)h, w = img.shape[:2]# 遍历每个像素点for i in range(h):for j in range(w):if mask[i, j] == 255.0:mask[i, j] = midbrightrate[i, j] = brightelse:mask[i, j] = (mid-1.0)/t*thresh[i, j]+1.0brightrate[i, j] = (1.0/t*thresh[i, j])*brightimg = img/255.0img = np.power(img, 1.0/mask[:, :, np.newaxis])*(1.0/(1.0-brightrate[:, :, np.newaxis]))img = np.clip(img, 0, 1.0)*255.0return img.astype(np.uint8)def lighting_fast(img, light):assert -100 <= light <= 100max_v = 4bright = (light/100.0)/max_vmid = 1.0+max_v*brightprint('bright: ', bright, 'mid: ', mid)gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY).astype(np.float32)/255.0thresh = gray*grayt = np.mean(thresh)# 使用numpy来计算可以加速，速度远快于上面的遍历mask = np.where(thresh > t, 255, 0).astype(np.float32)brightrate = np.where(mask == 255.0, bright, (1.0/t*thresh)*bright)mask = np.where(mask == 255.0, mid, (mid-1.0)/t*thresh+1.0)img = img/255.0img = np.power(img, 1.0/mask[:, :, np.newaxis])*(1.0/(1.0-brightrate[:, :, np.newaxis]))img = np.clip(img, 0, 1.0)*255.0return img.astype(np.uint8)if __name__ == '__main__':input_img = cv2.imread('tmp/302.png')light = 50start_time = time.time()res = lighting(input_img, light)print('time: {:.3f} s'.format(time.time() - start_time))cv2.imwrite('tmp/302_lighting_{}.jpg'.format(light), res)start_time = time.time()res = lighting_fast(input_img, light)print('fast time: {:.3f} s'.format(time.time() - start_time))cv2.imwrite('tmp/302_lighting_fast_{}.jpg'.format(light), res)

运行结果如下：

bright:  0.125 mid:  1.5
time: 6.454 s
bright:  0.125 mid:  1.5
fast time: 0.280 s

可以看到numpy加速很多倍，运行图的结果如下

原图：

 python版结果：

+50

-50：

 可以和c++原文对比是一致的