Research On Single Image Deraining Algorithm Based On Deep Learnin

Rain streaks can significantly occlude important objects in a scene,leading to the degradation of image content.Therefore,single-image deraining is an important research topic in the computer vision field.Currently,there are three main challenges in deraining tasks:Firstly,it is difficult to encode rain streaks in different directions.Secondly,due to the complexity of rain types,it is challenging to separate rain streaks from the image.Thirdly,many methods apply the same operations to all types of rain,which can reduce the model’s generalization ability.To address the issues mentioned above with single-image deraining,we propose a multi-stage and multi-scale attention fusion network and a scale constraint iterative update network.The key contributions and novel aspects of this paper are as follows:(1)A multi-stage and multi-scale attention fusion network is designed.It consists of two encoder-decoder networks,where the first stage network coarsens the features and the second stage network further refines them.The joint channel coordinate attention block accurately encodes rain streak features in different directions.The Inception attention branch block refines multi-scale features.The multi-level feature fusion block strengthens multi-scale feature representation.(2)A scale constraint iterative update network is designed.It includes a feature extraction block for accurately extracting rain streak features and image background details.The multiscale constraint block fuses contextual features at different scales from two sub-networks.Our proposed Global Gate Unit captures long-range dependencies and improves feature representation.The iterative update block optimizes deraining results at each iteration by using features from different scales.(3)A dynamic inference mechanism is designed to dynamically select the number of iterations based on different types of rain streaks,thereby improving the generalization ability of the model.(4)The proposed deraining algorithm achieves excellent performance on public datasets,demonstrating the effectiveness of our approach.