Multi-stage rain removal network based on union attention mechanism
In order to improve the quality of image shot in rainy environment,a Multi-stage Union Attention Network(MUANet)is proposed.The network has three stages.The initial input of each stage is processed the Union Attention Block(UAB)based on channel attention and spatial attention.UAB can detect the feature distribution in the channel and obtain the spatial information of rain patterns at the same time.In the first two stages,the encoding and decoding network with Half Instance Normalization Block(HINB)is used to mine the deep context information and accurately locate the position of rain texture in the image.In the last stage,with the attention map generated in the first two stages,the rain line texture is removed and the background details are restored.Experiments showed that MUANet exceeded than the existing methods in the performance of both rain removal and restoration of image background details.
multi-stage networkattention mechanismimage rain removalHalf Instance Normalizationencoding and decoding network
万方数据
维普
