ADR-Net: Attention-oriented detail recovery network for document image shadow removal

Fan Yang, Nanfeng Jiang, Da-Han Wang, Xu-Yao Zhang, Yun Wu, Shunzhi Zhu

Abstract: Existing methods based on deep learning have extensively explored the problem of document image shadow removal. However, most of them seldom consider the key regions of complex shadows and ignore detail preservation. Moreover, they usually have large model parameters that limit the potential values in real-world applications. To address these issues, we propose a simple but effective Attention-oriented Detail Recovery Network (ADR-Net) to remove complex shadows while preserving details in low complexity. In particular, on one hand, we explore the properties of shadows in color space and use luminance information to guide and generate shadow attention maps, which can accurately capture complex shadow distributions. For this purpose, we further design a Shadow Attention Generation Sub-Network (SAGN) that uses Multi-scale Large Kernel Attention (MLKA) mechanism to obtain long-range dependencies of shadows at various granularity levels. On the other hand, we propose a Dynamic Fusion (DF) strategy to avoid the ambiguity issues from wrong attention map during the learning process. In addition, we propose a Detail Refinement Sub-Network (DRN) that adopts Lightweight Spatial-Channel Convolution (LSCC) to facilitate recover details while decreasing redundant computing. Extensive experiments on public benchmarks and Optical Character Recognition (OCR) performance validate the effectiveness of our proposed ADR-Net and its superiority over state-of-the-art methods.