| Rainfall and snowfall severely degrade the accuracy of computer vision systems.In response to this problem,there has been widespread interest in how to effectively remove rain and snow traces while preserving as much of the original image information as possible.The thesis is based on a deep learning algorithm to investigate the single-image rain and snow removal problem,with the following main research elements:(1)Proposed a Transformer-based multi-stage continuous learning residual deraining network.This method is used for single image deraining work.The method captures the interaction of long-and short-term information by using the residual Fast Fourier Transform convolution module as the baseline of the deraining network,integrating high and low-frequency background information for transmission.The Contextual Transformer module is also introduced to focus on the rain streaks feature,using the contextual information of the rain streaks neighborhood to better recover the texture information in the image background,and using a multi-stage cyclic rain removal method to gradually remove the rain streaks.The method uses a continuous learning method,PIGWM,guided by the importance of the parameters,to improve the problem of catastrophic forgetting in deep learning by enabling the network model to have "memory" of the previous dataset.The thesis demonstrates the algorithm’s effectiveness on rain removal through comparative experiments,and ablation experiments prove that each module enhances the rain removal effect.(2)Proposed a single-image desnowing network based on multi-scale generative adversarial networks.The method uses the residual module as the backbone of the generator and reconstructs the appearance of the snow-free image generation a priori by introducing the Contextual Transformer module,focusing on the snow streaks features and using the contextual information of the snow streaks neighborhood to recover the background information obscured by snow marks.A learnable RN-L is also used to automatically detect potentially corrupted and undamaged regions for separate normalization and to perform a global affine transformation to enhance their fusion.The method uses a multi-scale discriminator on discriminator to make the discrimination more adequate and retain more details.The thesis demonstrates the algorithm’s effectiveness on snow removal through comparative experiments,while ablation experiments prove that each module enhances snow removal.(3)Produced a car license plate dataset under rain and snow,selected YOLOv5s+LPRNet as the license plate recognition network,and performed license plate recognition on the car license plate dataset containing rain and snow and the car license plate dataset after de-rain and de-snow with the de-rain and de-snow network proposed in the thesis,respectively,with the number of accurately recognized license plates as the evaluation index,and proved the effectiveness of the de-rain and de-snow network proposed in the thesis in the field of license plate recognition. |
