Multi-resolution Parallel Aggregation Network for Single Image Deraining

Abstract

AbstractSingle image deraining is an important preprocessing task, as rain streaks awfully reduce the image quality and hinder the subsequence outdoor multimedia issues. In this paper, we explore the multi-resolution representation for rain streaks through parallel hierarchical structure and multi-scale feature extraction and fusion, termed Multi-resolution Parallel Aggregation Network (MPA-Net) in end-to-end manner. Specially, considering the significant role of multi-resolution, we employ the first stage to capture the high-resolution features, progressively introduce high-to-low resolution streams to produce more stages, and then connect all stages in parallel. In each stage, Densely Connected Residual (DCR) block is involved to guide the feature extraction. Besides, Cross-Scale Feature Fusion (CSFF) is first introduced to receive and consolidate the correlated features from different scales followed with Squeeze-and-Excitation (SE) blocks, leading to rich the resolution representations. Extensive experiments demonstrate that our method outperforms the recent comparing approaches on the frequent-use synthetic and real-world datasets.KeywordsSingle image derainingMulti-resolution parallelCross-Scale Feature Fusion