Study On Image Deraining Based On Dual Encoders For Feature Extraction

When camera equipment is outdoors,it is inevitable that images containing the adverse effects of rain will need to be processed.The occlusion of rain streaks makes it difficult to distinguish detailed textures,which makes it difficult to collect photos and videos for applications such as object detection and instance segmentation in late stages.When computer vision systems are confronted with images containing these effects,their performance tends to degrade.Therefore,it is important to develop algorithms that can automatically remove these effects.The study of image deraining is very significant for computer vision research as it involves various fields including daily identity verification,intelligent transportation,video surveillance and autonomous driving.With the booming development of deep learning techniques and theories in the field of computer vision,significant breakthroughs have been made in deep learning-based image recovery algorithms.Such a deep learning model can automatically extract features in the washed-out rain-containing image supervised learning style based on a large amount of training data,and use it to learn a complex nonlinear mapping from the rain-containing image to the clean image.However,existing deep learning-based deraining models still have two major challenges:1.poor generalization ability of the trained obtained models;2.poor representation on images containing complex textures.To alleviate the impact of these two challenges,this paper makes the following work by combining two classical encoders in computer vision:(1)In this paper,we propose a neural network deraining algorithm using a dual encoder(convolutional neural network,Transformer)architecture to extract different features of a picture simultaneously.Two encoders are used simultaneously to extract features for an input rain-containing image,they are based on Transformer and convolutional neural network,respectively.CNN as a feature extractor can extract local features of the image,while transformer-type encoder can extract long-range semantic features.Since the two types of features do not belong to the same domain,a fusion approach is proposed in this paper to solve the semantic mutual exclusivity.By setting multiple variable scalar parameters that act on the features extracted by the Transformer encoder,then splicing them with the features of the CNN,and then flowing into a convolutional layer to achieve feature recombination after the splicing,and the effectiveness of the algorithm is demonstrated experimentally.(2)In this paper,we propose a multi-encoder neural net de-rain algorithm based on image edge information guided.Since rain streaks make the texture of complex images difficult to distinguish,this paper proposes a network structure based on image edge information to converge the edge features of rain-containing images into the neural net decoder to assist in improving the overall deraining performance of the algorithm.Specifically,a convolutional neural network based on channel attention and edge map is proposed to extract features of edge images,and the module is combined with the dual encoder neural network algorithm in Chapter Ⅲ to obtain a neural network algorithm with better deraining performance.Finally,the effectiveness of the module and the superiority of the proposed algorithm are experimentally demonstrated.In summary,the proposed method based on dual-encoder feature extraction has superior performance for the deraining task and can be used as an end-toend backbone network.In the experiments of Chapters 3 and 4,the quantitative metrics illustrate the better generalization capability of the proposed method,and the visual qualitative analysis illustrates that the proposed multi-coder structure design can better recover the detailed texture of images containing rain streaks.In particular,the proposed method can better identify rain streaks and recover high quality rain streak-free images when dealing with images containing single direction rain streaks and those with complex rain streaks.