Scholarlink

CNNs excel in computer vision with strong inductive biases and efficiency, but ineffective attention mechanisms can limit performance. Methods like SE-Block and CBAM often reduce channel dimensions or increase complexity, affecting accuracy and efficiency. We propose Fusion Attention (FA), a lightweight mechanism integrating channel and spatial attention to boost CNN performance. FA uses global average and max pooling with 1D convolutions for channel attention, and spatial pooling with 2D convolutions for spatial attention, fusing them additively without channel reduction. Experiments on PASCAL VOC show FA outperforms state-of-the-art methods in accuracy while maintaining efficiency, making it ideal for real-time tasks on resource-constrained devices.

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

Fusion Attention Efficiently Integrating Channel and Spatial Attention for CNN Enhancement

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wei Xu,
  • Zhonglin Ye,
  • Lin Li

摘要

CNNs excel in computer vision with strong inductive biases and efficiency, but ineffective attention mechanisms can limit performance. Methods like SE-Block and CBAM often reduce channel dimensions or increase complexity, affecting accuracy and efficiency. We propose Fusion Attention (FA), a lightweight mechanism integrating channel and spatial attention to boost CNN performance. FA uses global average and max pooling with 1D convolutions for channel attention, and spatial pooling with 2D convolutions for spatial attention, fusing them additively without channel reduction. Experiments on PASCAL VOC show FA outperforms state-of-the-art methods in accuracy while maintaining efficiency, making it ideal for real-time tasks on resource-constrained devices.