Research And Application Of Low-light Image Enhancement Algorithm Based On End-to-end Neural Networ

In low-light conditions,the color and detail information in images are usually significantly weakened,posing significant challenges to human vision and machine vision systems.In the field of machine vision,low-light images often exhibit problems such as low signal-to-noise ratios and lost texture details.To address these issues,low-light images can be enhanced to restore their quality to that of normal brightness images,providing support for subsequent tasks such as object detection,object tracking,and image classification.Currently,numerous deep learning-based image enhancement algorithms have been proposed for images in low-light environments.However,the performance of these algorithms is not ideal for images with severely insufficient exposure.To address these issues,this study employs deep learning methods to investigate an end-to-end neural network-based low-light image enhancement method.The specific work includes the following aspects:(1)A multi-scale low-light image enhancement network,which integrates dense residual blocks,was proposed to simultaneously address the difficulty of existing algorithms in brightness,contrast,color and detail.First,an input module is used to generate feature-rich input from images,which is then fed into a multi-scale main enhancement network integrated with residual dense blocks.The main enhancement network adopts a grid structure to fuse features at different scales.Finally,a refinement module is employed to enrich the image details and remove halo effects.Experimental results show that this method can simultaneously improve contrast and brightness,make colors more realistic,and enhance image texture details while effectively reducing noise and artifacts,offering a wide range of application prospects.(2)Fractal pyramid network with illumination information,a two-stage low-light image enhancement network,was designed to address the color distortion and detail loss problems caused by insufficient illumination estimation in traditional three-stage architecture networks.On one hand,a U-Net network with added spatial channel attention mechanisms is employed to extract illumination information to prevent uneven exposure and overexposure.On the other hand,a new encoder-decoder approach is constructed by combining fractal networks and pyramid networks,which has multiple information processing paths to fully exploit the relationships between contextual information and facilitate the exchange of features between different scales,thereby enriching image detail textures.The experimental results show the good versatility and practicality of the network in various application scenarios.