Research And Application Of Image Deraining Algorithm Based On Deep Learning

Single image deraining is a complex computer vision task that poses several challenges,which requires the computer to remove rainwater and restore the background information in the regions obscured by rain streaks.Existing deep learning-based rain removal algorithms have difficulty completely removing all rain streaks when dealing with complex rain streak distributions in images.In addition,most deep learning-based image deraining algorithms design the network as a "black box",and the specific functions of each module are difficult to explain.Although some researchers have attempted to enhance the interpretability of the network by combining traditional prior algorithms with convolutional neural networks,this method requires manual priors setting,which limits the rain removal effect of the network.In order to further improve the deraining effect and enhance the quality of image restoration,this paper summarizes the existing deraining methods and proposes two single-image deraining algorithms based on convolutional neural networks.(1)Multi-Scale and Multi-Stage Deraining NetworkTo address the problem that existing deraining networks have difficulty capturing the distribution information of rain streaks in images,a multi-stage multi-scale deraining network is designed.In order to extract effective information from the rainy image,a feature attention module is designed to efficiently utilize useful information between feature channels through attention mechanism,and applies it to local deraining encoder-decoder module and global deraining module.In addition,a semantic feature fusion module is introduced to merge the features of the encoder-decoder,and to pass rich semantic information to the next stage.Furthermore,an information transition module is designed between the stages of the network,which not only transitions shallow information to the next stage,but also provides supervision.Experimental results show that the proposed algorithm is competitive compared to other state-of-the-art deraining algorithms in Rain100 H,Rain100L,and Test100 datasets.(2)A Deep Unfolding Deraining Network based on Accelerate Proximal Gradient Descen AlgorithmA deep unfolding deraining network based on accelerate proximal gradient descent algorithm is proposed to address the issue of lack of interpretability in deep neural networks and the difficulty of manually assuming priors in traditional algorithms.The proposed rain-streak mapping module which corresponds to the gradient descent operator,using the basic module to simulate the rain kernel to reduce the complexity of manually computing priors.To enhance the correlation between features in different stages,a feature fusion module and a clear feature extraction module are designed.The feature fusion module is used to fuse the features extracted by the proximal deraining network in adjacent stages,while the clear feature extraction module is responsible for extracting the clear features of rain images and injecting them into the next stage.The results of the experiments indicate that the algorithm we proposed performs competitively when compared to other advanced rain removal methods on various datasets such as Rain100 H,Rain100L,Test100,and Test1200.In conclusion,algorithm 1 extracts rain streak features using a local-to-global approach to gradually remove global rain streaks.Additionally,it utilizes both the local deraining encoder-decoder module and the global deraining module to enhance the network’s feature learning ability.Algorithm 2 enhances the interpretability of the network by using accelerate proximal gradient descent and successfully recovers clear images that are closer to the ground truth images in terms of visual effect.The proposal of these algorithms provides reference for promoting research on single-image rain removal tasks in the future.