Scholarlink

<p>Shadow removal, a pivotal task in computer vision, significantly impacts subsequent processing tasks such as image segmentation and object detection. To achieve efficient shadow removal with minimal computational overhead and model size, this paper presents LSFormer, a lightweight network. LSFormer introduces the DiNA attention mechanism to model contextual relationships between shadow and non-shadow regions. Furthermore, it proposes a dual-scale gated enhancement feedforward network (DsGeFF) to effectively improve removal accuracy while reducing computational cost. To preserve shadow edges and texture details, a structure loss function is employed based on the Pearson correlation coefficient. Experimental results on the ISTD+ and SRD datasets show that, compared to ShadowFormer, LSFormer reduces the computational load from 63.1 GFLOPs to 20.76 GFLOPs and the number of parameters from 11.4M to 4.52M, while improving the PSNR on the ISTD+ dataset from 35.46 to 36.00. This method achieves a well-balanced trade-off between performance and efficiency, surpassing existing mainstream models and demonstrating strong practical applications. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient shadow removal using LSFormer with DiNA attention and dual-scale gated feedforward network

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Shuainan Hou,
  • Tao Han,
  • Shuaishuai Yu,
  • Jiale Pang,
  • Yourui Huang

摘要

Shadow removal, a pivotal task in computer vision, significantly impacts subsequent processing tasks such as image segmentation and object detection. To achieve efficient shadow removal with minimal computational overhead and model size, this paper presents LSFormer, a lightweight network. LSFormer introduces the DiNA attention mechanism to model contextual relationships between shadow and non-shadow regions. Furthermore, it proposes a dual-scale gated enhancement feedforward network (DsGeFF) to effectively improve removal accuracy while reducing computational cost. To preserve shadow edges and texture details, a structure loss function is employed based on the Pearson correlation coefficient. Experimental results on the ISTD+ and SRD datasets show that, compared to ShadowFormer, LSFormer reduces the computational load from 63.1 GFLOPs to 20.76 GFLOPs and the number of parameters from 11.4M to 4.52M, while improving the PSNR on the ISTD+ dataset from 35.46 to 36.00. This method achieves a well-balanced trade-off between performance and efficiency, surpassing existing mainstream models and demonstrating strong practical applications.