Low-light Image Enhancement And Super-Resolution Based On Dual Codecs

As a communication medium,images play a pivotal role in the transmission of information.In scenarios such as forest fire prevention and rescue and disaster relief,people need to shoot in the harsh field environment and transmit the images or videos in real time to the command center for playback on the big screen.However,limited by the environment,maybe only portable and low-quality filming equipment can be used.Images and videos taken at night and other low light conditions are not bright enough to see the objects clearly.Their resolution is also too low and would show severe blur when put on the big screen in the command center.This paper addresses above problems by innovatively improving the current mainstream low-light image enhancement algorithm.The main work consists of the following.1.A dual codec-based model for low-light image enhancement and super-resolution is proposed.The model applies the idea of multi-task learning,learning both low-light image enhancement and super-resolution simultaneously,using super-resolution as an auxiliary task for low-light image enhancement,helping the low-light image enhancement to better recover the brightness of the details in the images.This model adopts a dual codec structure:the light enhancement branch and the detail refinement branch.The parameters of the encoder part of two branches are shared to extract common features,and then the images are reconstructed by separate decoders.For multi-task learning,the loss functions corresponding to the tasks are designed.To address the problem of color distortion,which is common in low-light image enhancement algorithms,this paper introduces a color loss in the HSV color space.Experiments show that each module in the dual codec-based low-light image enhancement and super-resolution model contributes significantly to the overall performance.2.For multi-task learning,cross-task feature fusion is a very core part,and the performance of the feature fusion determines the performance of the algorithm.To address the problem of insufficient cross-task feature fusion,this paper introduces the channel attention mechanism on the basis of the dual codec-based low-light image enhancement and super-resolution model,improves the cross-task feature fusion approach,adds a feature fusion module,and uses the image detail features of the detail refinement branch to assist the learning of the light enhancement branch.Extensive experimental results show that the improved model achieves good results in terms of objective metrics,subjective visual effects and detail recovery,proving the effectiveness of the model and the improvements.