Low-light image enhancement based on GAN with attention mechanism and color Constancy

Images captured in low-light often suffer from severe quality degraded problems, such as low contrast and color distortion, which make it intractable for further computer vision tasks. To solve the problems above, we proposed a trainable parallel network including Brightness Enhancement Module based on GAN and Color Fidelity Module, which are guided by attention mechanism and color constancy respectively. The experimental results show that the proposed method could effectively improve the image contrast and preserve the color. The proposed method performs better than the state-of-the-art image enhancement methods (e. g. GAN based method) for improving the quantitative assessment including PSNR (19.72, up arrow 2.6%), BIQI (71.37, up arrow 1%), CIEDE2000 (4.97, down arrow 52%) and Pearson Correlation Coefficient (0.86, up arrow 105%).