Scholarlink

Large Kernel Attention (LKA) has demonstrated superior performance in image classification/segmentation and surpassing transformer-based visual tasks. However, using 1  \(\times \)  1 Conv for full-channel interaction in Large Kernel Separable Attention (LSKA) results in quadratic growth in computational complexity and memory usage as the number of channels increases. To alleviate this issue, this study proposes a Local-channel Large Separable Kernel Attention (LLSKA) module where the group convolution is employed to focus module solely on local-channel information to reduce the model complexity while maintaining system performance. The number of groups is determined by an adaptive strategy to determine the local-channel receptive ranges.

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

Local-Channel Large Separable Kernel Attention

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zengjie Deng,
  • Tao Lei,
  • Haixia Cui

摘要

Large Kernel Attention (LKA) has demonstrated superior performance in image classification/segmentation and surpassing transformer-based visual tasks. However, using 1  \(\times \)  1 Conv for full-channel interaction in Large Kernel Separable Attention (LSKA) results in quadratic growth in computational complexity and memory usage as the number of channels increases. To alleviate this issue, this study proposes a Local-channel Large Separable Kernel Attention (LLSKA) module where the group convolution is employed to focus module solely on local-channel information to reduce the model complexity while maintaining system performance. The number of groups is determined by an adaptive strategy to determine the local-channel receptive ranges.