zhaozunjin/retinexdip

The pytorch implementation of RetinexDIP, a unified zero-reference deep framework for low-light enhancement.

retinexdip's Issues

About NIQE

Hi, it's a good job. May I ask how to calculate NIQE？

增强的结果图没有用DIP网络得到的光照分量与反射分量相乘除以经过伽马校正的光照分量图，而是使用原图去除以光照分量得到最终的结果

   (R, G, B) = self.img.split()这个是原图

    ini_illumination = torch_to_np(self.illumination_out).transpose(1, 2, 0)
    # ini_illumination = misc.imresize(ini_illumination, (self.size[1], self.size[0]))
    ini_illumination = cv2.resize(ini_illumination, (self.size[0], self.size[1]))
    # print(ini_illumination.shape)
    ini_illumination = np.max(ini_illumination, axis=2)
    ini_illumination_renorm = (255 * ini_illumination).astype('uint8')
    # cv2.imwrite('output/illumination/illumination-{}.png'.format(step), ini_illumination_renorm)
    # If the input image is extremely dark, setting the flag as True can produce promising result.
    if flag==True:
        ini_illumination = np.clip(np.max(ini_illumination, axis=2), 0.0000002, 255)
    else:
        ini_illumination = np.clip(self.adjust_gammma(ini_illumination), 0.0000002, 255)
    R = R / ini_illumination
    G = G / ini_illumination
    B = B / ini_illumination
    self.best_result = np.clip(cv2.merge([B, G, R])*255, 0.02, 255).astype(np.uint8)
    cv2.imwrite('output/result-{}.png'.format(step), self.best_result)

The image is almost white

The code can run through after modification, but the image is almost white after enhancement

indicator test code

Dear zhaozunjin, I'm sorry to bother you, but I have something to ask you. It seems that there is no indicator test code in your code. Could you share the test code of NIQE, NIQMC and CPCQI with me? If you can send it to email: [email protected] 。 Thank you very much for your sharing.

在LOL数据集测试，没有得到好的结果，请问您超参数的设置