A Transformer-Driven Approach for Accurate and Efficient Single Image Deraining

Abstract--Single image deraining is of primary importance in the enhancement of computer vision tasks in bad weather conditions. Traditional CNN-based image deraining methods are generally bad at modeling long-range persistence. In this regard, transformer-based methods provide global contextual modeling but lack specialized attention mechanism to handle localized patterns of rain in their removal. For this, Transformer-Driven Image Deraining Network (TDIDNet) is proposed, that effectively combines vision transformers (ViTs) with a convolutional attention mechanism to remove rain streams while preserving fine details. TDIDNet consists of two main steps, Feature Extraction and Feature Aggregation & Refinement. The first step analyses the input images through three parallel convolution paths, using Channel Attention Blocks and typical convolutions before passing their features through the ViT modules. The second step consists of using Image Original Resolution Block (IORB) to restore structural integrity. TDIDNet achieves superior PSNR, SSIM, and loss scores, outperforming state-of-the-art transformer-based methods. KEYWORDS-Image Deraining, Transformer-Driven Image Deraining Network (TDIDNet), vision transformer (ViT), Attention Module (AM), Channel Attention Block (CAB), Image Original Resolution Block (IORB).