关键词: 低照度偏振图像, 双分支特征提取, 残差空洞卷积, 边缘结构相似性损失 

Abstract:  In order to solve the problems of low brightness, serious noise, and color distortion of color polarization images in low-illumination scenes, an unsupervised learning algorithm for color enhancement of low- illumination color polarization images is proposed, which is named LPEGAN(Low-Light Polarization Enhance Generative Adversarial Network). Firstly, a double-branch feature extraction module is designed and used different branches to extract features from Stokes parameters S0 and S1 ,S2 , respectively. Secondly, the residual void convolution module is constructed. And the different expansion rates can expand the receptive field to improve the model extraction ability and reduce the image color distortion. The edge texture loss function is constructed to ensure the structural similarity between the enhanced image and the input image. Experimental verification is carried out on the public datasets LLCP(Low-Light Chromatic Intensity-Polarization Imaging), IPLNet(Intensity-Polarization Imaging in Low Light Network), and self-built datasets. The experimental results show that the proposed algorithm has better visual effects, and all evaluation indicators are significantly improved. Polarized image brightness is enhanced, noise is significantly suppressed, and image colors are more realistic and natural.

Key words: low-light polarized image, double-branch feature extraction, residual void convolution, edge structural loss