Adjustable enhancer for low-light image enhancement using multi-expressions fusion and convolutional kernel calibration

Owing to constant convolutional kernels in trained network, conventional enhancer for low-light image can only generate enhanced image with an invariable level of brightness. To address this issue, we propose a novel enhancer based on multi-expressions fusion and kernel calibration, which can improve visibility, remove noise, and flexibly adjust the brightness of enhanced image. In this study, a strategy of fusing multiple expressions is incorporated into the enhancer to improve the performance in terms of representation learning. Subsequently, a kernel calibration network is designed to generate dynamic kernel calibration maps. With these dynamic calibration maps combining with static convolutional kernels, dynamic convolution and adjustable brightness are incapable of being achieved. Furthermore, a method of synthesizing training data is employed for generating simulated low-light images to remove noise while enhancing low-light image. Comprehensive experiments validate the exceptional performance of the proposed enhancer in improving the visual quality of low-light image and flexibly adjusting the brightness of enhanced image.