Low-light/Haze Image Enhancement And Restoration Based On Deep Learning

Image restoration is an important research content in the field of image processing and computer vision.Under conditions such as low light and haze,the images collected by outdoor computer vision systems usually suffer severe degradation,such as low brightness,low contrast,and loss of detail,which seriously affect its application in outdoor computer vision systems.Therefore,it is of great significance to enhance and restore images collected by outdoor computer vision systems under low-light,haze and other conditions.This paper focuses on the research of deep learning methods for the restoration and enhancement of low-light/haze images.The main work and the achievements are as follows:(1)In terms of the problems that traditional low-light image enhancement methods are sensitive to noise and poor generalization of deep learning methods,an attention-adaptive convolutional neural network for low-illuminance image enhancement was proposed.It consists of three subnets:a gradient restoration subnet based on the encoder-decoder network,a brightness restoration subnet based on the encoder-decoder network,and a backbone network for adaptive adjustment of illumination and contrast.In order to avoid weakening the contrast enhancement when restoring the image brightness,a new adaptive instance normalization module with linear contrast enhancement capability is constructed in the backbone network of the brightness and contrast adaptive adjustment.At the same time,the integrity of the enhanced image structure and texture is ensured by adding gradient restoration subnet.Experimental results and quantitative analysis show that compared with the existing methods,the method in this paper can achieve better enhancement effects,and it can also achieve better enhancement effects for complex low-light scenes with high-light areas and underwater low-light scenes with uneven lighting.Related work has been submitted to "IEEE Transaction Image Processing"(under review).(2)In view of the problems that the existing image dehazing methods cannot recover haze images with large sky areas or heavy haze images,a context-guided generative adversarial network for image dehazing was proposed.In this method,the generator adopts a new encoder-decoder network composed of feature extraction subnet,context extraction subnet and fusion subnet.The feature extraction subnet is an encoder used to extract features of haze images;the context extraction subnet is a multi-scale parallel pyramid decoder used to extract deep features of the encoder and generate coarse dehazing images;the fusion subnet is a decoder that combines feature extraction network features and context extraction network features and performs further restoration.In order to obtain a better dehazing effect,the multi-scale information obtained in the decoding process of the context extraction subnet is used as guidance information to optimize the learning process of the fusion subnet.In order to ensure that the proposed network can work effectively under different haze conditions,the two decoders adopt different loss functions.The experimental results show that this method has better dehazing performance compared with the existing methods.Related work has been officially published in "IET Image Processing"(CGGAN:A Context Guided Generative Adversarial Network For Single Image Dehazing,IET Image Processing,14(15):3982-3988,2021.).(3)In view of the problem of clearing and restoring night haze images,a generative adversarial network based on attention fusion was proposed.The network consists of a low-light image enhancement network based on convolution,an image dehazing network based on convolution,a generate network based on multi-head input and attention fusion,and a discriminant network based on convolution.The method idea is:first pre-training the low-light image enhancement network and the image dehazing network,and then use the output image of the pre-trained low-light enhancement network and the pre-trained image dehazing network as the input image together with the night haze image to be entered into the generator.Finally,the night haze image is further enhanced by the method of attention fusion and generative adversarial.The experimental results show that the method in this paper successfully adds the low-light image enhancement method and the image dehazing method to the restoration task of the night haze image,and further improves the quality of the restoration result.The related work has been completed.