Scholarlink

<p>One of the most widely applied tasks in security and computer vision applications today is object detection, by which different categories of objects can be detected and located from images and videos. The advances in Deep Learning (DL) have presented us with new ways to detect objects and develop efficient mechanisms for object detection. The state-of-the-art research field shows that different activation functions in different layers and structures of the DL network significantly impact accuracy; however, they are mainly focused on selecting the most suitable activation functions for different layers from a local optimization perspective. In this work, we investigate the combination of activation functions across different layers from a global perspective as an optimization problem, applying orthogonal experiments to optimize the combined activation functions. Also, we design orthogonal matrices using different layers in the object detection deep learning model and several standard activation functions. We achieved a Mean Average Precision (mAP) of 83.34% on the PASCAL VOC 2007 dataset and 84.62% on the PASCAL VOC 2012 dataset. Experiments on two classical object detection datasets demonstrate that globally optimized results show significant improvements with statistical significance. The code has been made publicly available at <a href="https://github.com/54yc/OCA.">https://github.com/54yc/OCA.</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

A method for composite activation functions in deep learning for object detection

  • 期刊论文

  • 发表日期：

    2025年3月6日

查看全文
  • Jing Liao,
  • Chang Yu,
  • Lei Jiang,
  • Linpei Guo,
  • Wei Liang,
  • Kuanching Li,
  • Al-Sakib Khan Pathan

摘要

One of the most widely applied tasks in security and computer vision applications today is object detection, by which different categories of objects can be detected and located from images and videos. The advances in Deep Learning (DL) have presented us with new ways to detect objects and develop efficient mechanisms for object detection. The state-of-the-art research field shows that different activation functions in different layers and structures of the DL network significantly impact accuracy; however, they are mainly focused on selecting the most suitable activation functions for different layers from a local optimization perspective. In this work, we investigate the combination of activation functions across different layers from a global perspective as an optimization problem, applying orthogonal experiments to optimize the combined activation functions. Also, we design orthogonal matrices using different layers in the object detection deep learning model and several standard activation functions. We achieved a Mean Average Precision (mAP) of 83.34% on the PASCAL VOC 2007 dataset and 84.62% on the PASCAL VOC 2012 dataset. Experiments on two classical object detection datasets demonstrate that globally optimized results show significant improvements with statistical significance. The code has been made publicly available at https://github.com/54yc/OCA.