Adversarial Context Aggregation Network for Low-Light Image Enhancement

Abstract

Image captured in the low-light environments usually suffers from the low dynamic ranges and noise which degrade the quality of the image. Recently, convolutional neural network (CNN) has been employed for low-light image enhancement to simultaneously perform the brightness enhancement and noise removal. Although conventional CNN based techniques exhibit superior performance compared to traditional non-CNN based methods, they often produce the image with visual artifacts due to the small receptive field in their network. In order to cope with this problem, we propose an adversarial context aggregation network (ACA-net) for low-light image enhancement, which effectively aggregates the global context via full-resolution intermediate layers. In the proposed method, we first increase the brightness of a low-light image using the two different gamma correction functions and then feed the brightened images to CNN to obtain the enhanced image. To this end, we train ACA network using L1 pixel-wise reconstruction loss and adversarial loss which encourages the network to generate a natural image. Experimental results show that the proposed method achieves state-of-the-art results in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM).1