Single image deraining using asymmetric feature pyramid context-aware network

Abstract

Existing image deraining algorithms have challenges of accurately identifying the size and density of rain streaks, which can lead to incomplete removal and difficulties in restoring high resolution images. To address these challenges, we propose a context-aware network based on an asymmetric feature pyramid (CA-AFPN) for effective rain streak removal. The CA-AFPN is composed of feature extraction module, image restoration module, and a multi-scale feature fusion module. In the feature extraction module, the key features are extracted using a channel and self-attention (CS-Attention) module, which can perform downsampling on the image to capture color, semantic, and spatial information from various feature layers. The image restoration module employs the context-aware deep upsampling (CADU) technique to globally and dynamically restore the original features. Additionally, horizontal connections between the two modules integrate shallow physical positioning and deep semantic information, expanding the network receptive field and spatial context. Finally, the multi-scale feature fusion module (MFF) utilizes a residual network and dilated convolution layers to merge feature information across different scales for reconstructing the rain-free image. Extensive experimental results demonstrate that the proposed method is effective not only on synthetic datasets but also achieves superior performance on real-world data.