Research On Single Image Deraining And Lightweighting Based On Deep Learnin

The large number of photographs captured by cameras are often affected by meteorological factors,especially rainfall,leading to the appearance of white stripes of varying sizes in the images,which can degrade image quality and impair the performance of outdoor computer vision systems.Deep learning methods have made significant progress in image deraining tasks by designing complex algorithms to learn the mapping relationship between rain streaks and clean backgrounds,achieving better visual effects.However,there are still two key challenges with current deep learning de-raining networks: first,the dataset is usually synthesized using software,making it difficult to cover all scenarios and resulting in a significant decrease in performance when applied to other datasets.Second,existing methods prioritize objective metrics and incorporate a large number of iterations,cycles,and multi-scale structures,leading to an increase in parameter volume and runtime,and poor real-time performance.To address these challenges,this paper proposes a lightweight single-image de-raining network with strong generalization ability that combines deep learning algorithms with prior domain knowledge.The main contributions are as follows:(1)A single-image rain removal model based on residual channel decomposition is developed to remove rain streaks and restore the complete background structure.The algorithm filters out rain streaks using residual channel decomposition,and an improved feature attention fusion module is used to interactively fuse the original image and the background features to avoid information loss and feature interference.Finally,the processed image is input into a semi-supervised deep learning network to restore high-quality rain-free images with accurate and clear structures.Experimental results on synthetic and real datasets show that the proposed model enhances the generalization ability of the rain removal network and alleviates the overfitting problem.(2)A novel gradient-based pruning method,named generalized through estimator,is proposed to maintain the clarity and quality of rain-removed images by the network after high compression.The method retains weight parameters with larger magnitudes and sets weight parameters with smaller magnitudes to zero during backpropagation,which restores activation states and preserves more important weight parameters.In addition,a soft-thresholding operator is used to reduce sharp changes in forward path weights and prevent layer collapse caused by hard-thresholding operator zeroing out weights below a certain threshold.Experimental results show that the proposed method achieves a good balance between running speed and image quality,and the image inference speed of the rain removal network with 90% sparsity is approximately 2.24 times faster than the original network.