多尺度注意力密集连接网络的单幅图像去雨算法

Single-image deraining algorithm of multi-scale attention dense connection network

为了获取更好的去雨效果,更好地保留图像纹理细节,设计了一个端到端的多尺度注意力密集连接网络,引入了多尺度注意力模块,通过不同的扩张卷积获得不同尺度下的雨条纹特征。设计了一个并行注意力蒸馏模块,结合通道注意力和子空间注意力机制来校准从多尺度采集获得的图像特征,然后消除冗余特征信息。整体网络架构采用密集连接跨层融合代替直接连接的方式,加强了特征重用,易于训练。结果表明,本方法在提升去雨性能的同时,改善了对于图像的细节保留。

In order to achieve better deraining effect and retain the texture details of images,an end-to-end multi-scale attention dense connection network was designed.A multi-scale attention module was introduced to obtain rain streak features at different scales through different expansion convolutions.A parallel attention distillation module was designed to calibrate image features obtained from multi-scale acquisition by combining channel attention and subspace attention mechanism,and then redundant feature information were eliminated.The overall network architecture adopts dense connection cross-layer fusion instead of direct connection,thus strengthening feature reuse and making it easy to train.The results show that the proposed method improves the image detail retention while improving the deraining performance.