Scholarlink

<p>In this paper, we introduce a novel lightweight transformer-based network, LDFormer, specifically designed for single image dehazing. LDFormer employs encoder-decoder architecture to efficiently restore haze-free images while tackling the critical challenge of computational complexity. To optimize feature transfer efficiency, we incorporated a selective feedforward network (SFN) within each transformer module (TM), which significantly reduces the computational burden by selectively processing relevant features, while maintaining high performance in dehazing tasks. Moreover, to improve the attention process in LDFormer, we introduced a cross-layer attentive integrated module (CAIM). Through the refinement of the network's emphasis on critical data, CAIM facilitates more effective feature extraction and usage at different transformer module levels. This combination enhances the network’s ability to capture detailed feature representations, leading to superior dehazing quality. Comprehensive tests conducted on multiple benchmark datasets confirm the effectiveness of our network. The results demonstrate that LDFormer not only surpasses current state-of-the-art (SOTA) techniques in terms of restoration quality but also offers significantly improved computational efficiency. Our network emerges as a robust solution for single image dehazing, excelling in both quantitative measurements and visual clarity, thus proving to be a highly effective and efficient choice for real-world dehazing 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

LDFormer: lightweight dehazing transformer

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • D. Pushpalatha,
  • P. Prithvi

摘要

In this paper, we introduce a novel lightweight transformer-based network, LDFormer, specifically designed for single image dehazing. LDFormer employs encoder-decoder architecture to efficiently restore haze-free images while tackling the critical challenge of computational complexity. To optimize feature transfer efficiency, we incorporated a selective feedforward network (SFN) within each transformer module (TM), which significantly reduces the computational burden by selectively processing relevant features, while maintaining high performance in dehazing tasks. Moreover, to improve the attention process in LDFormer, we introduced a cross-layer attentive integrated module (CAIM). Through the refinement of the network's emphasis on critical data, CAIM facilitates more effective feature extraction and usage at different transformer module levels. This combination enhances the network’s ability to capture detailed feature representations, leading to superior dehazing quality. Comprehensive tests conducted on multiple benchmark datasets confirm the effectiveness of our network. The results demonstrate that LDFormer not only surpasses current state-of-the-art (SOTA) techniques in terms of restoration quality but also offers significantly improved computational efficiency. Our network emerges as a robust solution for single image dehazing, excelling in both quantitative measurements and visual clarity, thus proving to be a highly effective and efficient choice for real-world dehazing applications.