Low-Light Image Enhancement Based on Joint Generative Adversarial Network and Image Quality Assessment

Abstract

Images captured in low-light conditions are often disturbed by low-light, blur and noise. Most of the conventional image enhancement methods are less robust without considering the effectiveness of the blur and noise. To enhance image equality under the complex environment, we propose a novel image enhancement method based on joint generative adversarial network (GAN) and image quality assessment (IQA) techniques. GAN can be well used for image enhancement in low-light case, but it is not robust in blur and noise case. IQA method uses CNN to evaluate each enhanced image quality based on some scores that correlates well with the human perception. The scores can guide the GAN learning for further enhancing the image quality. Instead of l2-term loss function, we define a multi-term loss function for its minimization to create a good image estimate. Experimental results demonstrate the proposed method is more effective than current state-of-art methods in terms of the quantitative and qualitative evaluation.