Research On Low-light Image Enhancement Based On Deep Network

Aiming at improving the visibility of an image,low-light enhancement is an important post-processing task.Based on the analysis about the causes of low light,the fundamental task of low-light image enhancement is defined.On this basis,we analyze the inadequacies of existing models,and propose appropriate algorithm to solve them.On one hand,the enhancement model based on autoencoder can well estimate the new illumination distribution in the imaging scene,but serious loss of details in the reconstruction process.Herein,to achieve improved results by combining the relevant conventional filtering techniques and deep learning model.On the other,although the Retinex Net model can enhance the contrast,additional noise is introduced in the process,resulting in an obvious black border effect,and the image is accompanied by color distortion.In this,the related problems are improved through the optimization of the loss function and network structure.1)A low-light image enhancement model based on base-detail decomposition was proposed.The enhancement model based on autoencoder shows its effectiveness in improving the visibility,but it has the limitation of losing fine-scale details during the decoder stage.To address this issue,we process the image layer only containing largescale image structure while remain the image details at fine scales,based on the observation that the primary goal of low-light enhancement is to improve the illumination distribution rather than image details.To this end,we first decompose an image into the base layer and the detail layer based on an edge-preserving filter.We learn an autoencoder network for reshaping the image illumination only based on the base layer.The detail layer is kept unchanged during the enhancement,and added back to recompose the final enhanced image.Qualitative and quantitative experiments were conducted to validate the effectiveness of the proposed method.Compared with traditional autoencoder model,our model is able to improve the visibility and preserve the image details simultaneously.2)A contrast enhancement model of convolutional neural network based on Retinex was proposed.Considering the problem of over-enhancement and distortion of the Retinex Net model in the edge contour and the reconstruction process,the network loss function and structure are improved.Based on the Retinex Net model,the loss function was redesigned.At the same time,some constraints are added to make the reflection component smoother and reduce the black edge effect.In addition,the decomposition module Decom-Net of Retinex Net model was replaced with the more advantageous technology U-Net,in order to get better illumination and reflection components.Compared with the Retinex Net,our model not only effectively suppressed the problems of edge contour and color distortion under the premise of effectively retaining the texture details of the image,but also improved the contrast of the image well.