A Low-Light Image Enhancement Algorithm Research Based On Image Fusion

The images or videos captured by sensor always have low contrast,poor image quality and lose numerous detailed information because of lacking of light.To improve the visual effect of low-light images so as to apply the images to the field of target recognition,it is necessary to use the image enhancement algorithms to enhance the details of the image and improve the image quality.However,the most of traditional low-light image enhancement algorithms cannot satisfy the expected requirements.Thus,this paper focuses on the research of low-light image enhancement algorithms,and makes improvement on existing algorithms.The main research results and content are as follows:1.Based on the image fusion,this paper presents a low-light image enhancement algorithm.The presented algorithm first generates the three enhanced images as input images through three image enhancement algorithms.Then,the three weighting maps corresponding to the three input images are calculated according to the effective information and features of three input images.Finally,the final enhanced image is obtained through weighting fusion of three input images and corresponding weighting maps.Experimental results show that the presented algorithm can enhance the brightness of image and keep the contrast of image well.2.This paper presents a low-light image enhancement algorithm based on deep residual network.Inspired by the idea of the fusion algorithm,we use the deep learning to learn the non-linear mapping model from the low-light image with poor brightness to the enhanced image with good brightness,without designing features and adjusting parameters artificially.Concretely,we use a deep residual network to learning a system that is suitable for low-light image enhancement.The simulation experiments show that the system can enhance the brightness of the low-light image well while maintaining the natural image.3.This paper establishes a low-light image database.In order to achieve supervised training of residual network,a large amount of training images are required.Thus,we use three degenerate models to degenerate the brightness of images,and generate the corresponding brightness degradation images.The brightness degradation images and their corresponding reference images constitute our training set.Through the experiment,it is clear that the residual network trained by the simulated image database can effectively enhance the low-light image.