Scholarlink

<p>Attention mechanisms are pivotal for enhancing deep convolutional neural networks (CNNs) in computer vision, yet existing methods often inadequately leverage spatial information, reduce channel dimensions, or escalate network complexity—challenges exacerbated by Transformer-like quadratic costs. This paper introduces efficient localization attention (ELA), a novel, lightweight module that achieves highly competitive performance among lightweight attention mechanisms across image classification, object detection, and semantic segmentation, while maintaining computational efficiency. By analyzing coordinate attention (CA), we identify its limitations, including batch normalization’s generalization issues, channel dimensionality reduction’s negative impact, and complex attention generation. To overcome these, ELA employs 1D convolutions and group normalization (GN) to encode spatial features without channel reduction, capturing long-range dependencies and position-sensitive information effectively. We developed four versions of ELA-T, ELA-B, ELA-S, and ELA-L—optimized for different visual tasks. Extensive experiments on ImageNet, MS COCO, and Pascal VOC demonstrate ELA’s superiority, achieving up to 2.4% Top-1 accuracy gain, 1.96% mAP improvement, and 1.08% Mean IoU increase over state-of-the-art methods, making it ideal for resource-constrained, real-world applications. Project page: <a href="https://github.com/Xuwei86/ELA_Code">https://github.com/Xuwei86/ELA_Code</a></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

ELA: efficient location attention for deep convolution neural networks

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Wei Xu,
  • Yi Wan,
  • Weina Zhao

摘要

Attention mechanisms are pivotal for enhancing deep convolutional neural networks (CNNs) in computer vision, yet existing methods often inadequately leverage spatial information, reduce channel dimensions, or escalate network complexity—challenges exacerbated by Transformer-like quadratic costs. This paper introduces efficient localization attention (ELA), a novel, lightweight module that achieves highly competitive performance among lightweight attention mechanisms across image classification, object detection, and semantic segmentation, while maintaining computational efficiency. By analyzing coordinate attention (CA), we identify its limitations, including batch normalization’s generalization issues, channel dimensionality reduction’s negative impact, and complex attention generation. To overcome these, ELA employs 1D convolutions and group normalization (GN) to encode spatial features without channel reduction, capturing long-range dependencies and position-sensitive information effectively. We developed four versions of ELA-T, ELA-B, ELA-S, and ELA-L—optimized for different visual tasks. Extensive experiments on ImageNet, MS COCO, and Pascal VOC demonstrate ELA’s superiority, achieving up to 2.4% Top-1 accuracy gain, 1.96% mAP improvement, and 1.08% Mean IoU increase over state-of-the-art methods, making it ideal for resource-constrained, real-world applications. Project page: https://github.com/Xuwei86/ELA_Code