Research On Low-Illumination Image Enhancement And Super-Resolution Reconstruction Algorithms

Affected rby natural environments such as rain,fog,and haze,or the imaging system of the acquisition device with low precision,and the image has low brightness,obvious noise,lack of detail,or blurred resolution,which makes it impossible for people to directly obtain useful information in image.Therefore,a low-illumination image enhancement algorithm based on WGIF(Weighted Guided Image Filtering),and a super-resolution image reconstruction algorithm based on depth residual learning are proposed in this paper.The main research work and results are as follows:(1)In low illumination environment,the image has low brightness,poor contrast,and difficult to distinguish the objects,etc.Therefore,based on the research of traditional Retinex algorithm,this paper proposes a low-illumination image enhancement algorithm based on WGIF.In proposed algorithm,WGIF is adopted twice.Among them,one is used to estimate the illumination component of image,and the obtained result is improved by local illumination enhancement methods,which overcomes the flaws of Retinex algorithm such as halo,blurred details.While the other is mainly used to process the reflection component,and the noise amplification phenomenon is effectively prevented.Simultaneously,the image brightness is adjusted by S-type hyperbolic tangent function,so the image contrast is enhanced to a certain extent.In addition,color space conversion and linear color restoration methods are used to process color images,effectively avoiding the color distortion and imbalance caused by the Retinex algorithm.Experimental results show that the proposed algorithm has good enhancement effect on the overall brightness,detail clarity and contrast of the image,and makes the color of the image more natural and visually better.(2)Both SRCNN and FSRCNN are image reconstruction models based on CNN(Convolutional Neural Network),but these models are relatively shallow in network depth,and the learned features are less,which leads to the decline in quality of image reconstruction.Therefore,this paper proposes a super-resolution image reconstruction model based on depth residual learning.Combined with the characteristics of SRCNN and FSRCNN,the depth of the model is extended to 13 layers and the residual learning structure is introduced in this model,which not only increase the quantity of image features,but also effectively alleviates the problem of training "degradation" caused by too many layers of neural network.Furthermore,deconvolution is the end of the model as the image upsampling operation,PReLU acts as the activation function after the convolution operation,and Adam method is the optimization strategy in the model training to update the network parameters,thus making the model training more robust and efficient.Experimental results show that the proposed reconstruction model in this paper is superior to several classical reconstruction methods in both qualitative and quantitative evaluations,and has better reconstruction effect for medical and depth images,especially for images with obvious texture and edge details or with strong contrast.