Research On Image Deraining Based On Transformer And Wavelet Analysis

Rain streaks in rainy weather can obscure and distort background information,causing a decrease in image quality and affecting downstream visual algorithms such as object detection and semantic segmentation.Therefore,the study of image deraining holds significant importance.However,many current deraining methods have the issue of insufficient detail restoration ability.This is because the majority of deraining methods primarily rely on convolutional neural networks,which have the limitations of small receptive fields and fixed weights,making it difficult for the network to separate rain streaks interweaved with the background.The recently emerged Transformer can address these limitations.Additionally,existing deraining methods have shortcomings in utilizing frequency-domain information and need to improve their modeling capabilities for different frequency-domain features.Based on the above analysis,the research work in this thesis is as follows:(1)To leverage the respective strengths of Transformer and convolutional neural networks and fully utilize frequency-domain information to solve the problem of insufficient detail restoration ability in existing deraining methods,we propose a Wavelet Domain Deraining Network based on Transformer Parallel Attention(WDDNet).WDDNet decomposes multiscale feature maps into low-frequency and high-frequency components using wavelet transform.It utilizes Multi-attention Blocks to learn background contours and texture details separately from the low-frequency and high-frequency components,enabling the blocks to focus on extracting specific information.The Multi-attention Block consists of the Parallel Attention Block and the Channel Group Attention Block.The Parallel Attention Block combines the nonlocal modeling ability of Transformer and the local modeling ability of convolution.The Channel Group Attention Block performs image-level interaction in the channel dimension,achieving spatial global attention.Experimental results show that WDDNet improves deraining performance by enhancing detail restoration ability,outperforming current advanced deraining methods in synthetic datasets,real datasets,and joint object detection tasks.(2)Due to the high complexity of rain streaks,single-stage networks often cannot completely remove rain streaks.Therefore,we further propose a Multi-stage Deraining Network based on Wavelet Convolutional Transformer(MSDNet).The Wavelet Convolutional Transformer enables self-attention to be conducted separately in the low-frequency and highfrequency domains,enhancing the model’s modeling ability for different frequency-domain features.Meanwhile,the convolution extracts local features across windows and in the feedforward layers,strengthening the local modeling ability of Transformer.The multi-stage strategy decomposes the deraining process into multiple sub-stages,allowing each sub-stage to focus on removing rain streaks of different levels and enhancing the network’s ability to detect complex rain streaks by fusing multi-scale features from different stages.Experimental results show that MSDNet has stronger deraining performance,outperforming current advanced deraining methods in synthetic datasets,real datasets,and joint semantic segmentation tasks.