RIADNet: Single Image Deraining Network for Raindrops and Rain Streaks Removal

Abstract

Both raindrops and rain streaks are common degradations in images captured on rainy days, which not only reduce the quality and visibility of images but also significantly affect downstream tasks such as object detection. However, most existing derain algorithms focus on one specific degradation, failing to provide a comprehensive analysis across other scenarios. In this paper, we propose RIADNet, a Rain Information Attention Deraining Network, which jointly removes raindrops and rain streaks while preserving critical image details. Initially, we devise a simple and efficient rain information attention module(RIAM) to extract raindrop and rain streak information from images accurately, guiding the network to focus on rainy regions and enhancing deraining performance. Furthermore, a multi-scale dilated convolution feature fusion module(MDFFM) integrates encoder features from multiple receptive fields through parallel dilated convolutions with varying dilation rates, which significantly improves multi-scale feature representation. Moreover, a deformable wavelet sampling module(DWSM) replaces traditional sampling with deformable wavelet-based kernels, adaptively preserving high-frequency details during feature extraction and minimizing information loss. Qualitative and quantitative experimental results on three public datasets demonstrate the superior performance of RIADNet in addressing diverse rain degradations. Notably, on the RDS dataset (mixed raindrops and rain streaks), RIADNet achieves a PSNR of 29.78 dB and SSIM of 0.921, outperforming all compared state-of-the-art deraining models while reducing parameters by 54.52% versus the second-best method.