Research On Low-light Image Enhancement Algorithms For Vision-like Processing

Visual perception,which is a major means for human beings to capture external information,can map the scene content into the perceptual content in the brain systems that make the people understand the image content and make decisions.Even in the dark scenes,the vision system still has strong recognition ability.Therefore,it is beneficial to efficiently enhance the brightness of low-light images by using the information processing mechanism theory of biological visual perception system and studying its characteristics to establish the deep neural network model.Inspired by this,this paper proposes low-light image enhancement algorithm based on visual processing,which mainly analyzes the information processing system of human retina in detail to design the model.This work consists of two parts.One is to propose the low-light image enhancement algorithm based on improved Retinex.The other is to design the progressive low-light image enhancement network similar to visual transduction.So,the main research contents of this work can be summarized as follows:(1)low-light image enhancement algorithm based on improved Retinex.Inspired by the theory of retinal cortex,low-light image enhancement algorithm based on improved Retinex is proposed.With the basic of Retinex-Net,firstly,the shallow upper and lower sampling structure is designed to decompose the dark light images into reflection component and illumination component that fully considers the color constancy of the retina to the object.Besides the total variational loss is used to smooth the gradient of reflection component.Secondly,in order to reduce the response of nonbrightness features and color distortion,the enhancement network uses the attention mechanism to improve the image illumination and constrains the low-frequency component of the image through color loss.The experiments show that compared with the benchmark algorithm,SSIM and PSNR are increased by 28.80% and 10.46%respectively,indicating that this algorithm can improve the problems of high noise and color distortion in Retinex net to some extent.(2)The progressive low-light image enhancement network similar to visual transduction.While enhancing the brightness,the low-light image enhancement task needs to consider other degradation problems,such as noise amplification,detail blur and so on.However,there are few algorithms to deal with these degradation problems effectively.So,based on this,inspired by the structure of human retina and the working mechanism of visual transduction pathway,the progressive low-light image enhancement network similar to visual transduction is designed.It composes of three branches: brightness enhancement,noise suppression and detail reconstruction,which simulates the serial processing mechanism of visual perception system to process the image step by step.In addition,visual enhancement block is proposed as the basic element to build the whole network,which includes three-level information processing path and adaptive selection unit.The experiments show that this network can effectively balance the problems of illumination enhancement,noise suppression and structure restoration,which is due to the structure of simulated visual information transmission pathway and the characteristics of biological visual attention selection.In summary,these papers mainly study low-light image enhancement algorithm based on visual processing,which includes two major tasks: the low-light image enhancement algorithm based on improved Retinex and the progressive low-light image enhancement network similar to visual transduction.The experimental results of these two works show that the algorithm can handle a variety of degradation problems of dark light images.More importantly,the model design principle follows the working mechanism of human visual system,which will promote the development of similar visual system and make new breakthroughs.