Scholarlink

<p>This paper presents a multi-stage deep learning method for detecting soccer events, built around three main modules. The first, the No-Highlight Detection Module, uses a Variational Autoencoder (VAE) trained only on highlight frames to filter out irrelevant or no-highlight scenes with higher precision than standard autoencoders. The second, the Image Classification Module, is an EfficientNet-based network trained on the newly created Soccer Event (SEV) dataset, which we collected and manually labeled from broadcast matches. The dataset contains 60,000 balanced images across ten soccer event categories. The third, the Fine-Grain Classification Module, employs OSME attention to distinguish between yellow and red cards, which are often visually similar. The SEV dataset is complemented by an independent evaluation set to assess generalization on unseen data. Experiments conducted on 50 full soccer matches from five different leagues, including the UEFA Champions League (UCL), English Premier League, Bundesliga, Brazilian Serie A, and the Iranian Pro League, demonstrate that the proposed system achieves competitive event detection accuracy while maintaining real-time performance. These results suggest that combining open-set filtering with fine-grain attention enhances precision and robustness under realistic broadcast conditions.</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

Soccer event detection using deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • Ali Karimi,
  • Ramin Toosi,
  • Mohammad Ali Akhaee

摘要

This paper presents a multi-stage deep learning method for detecting soccer events, built around three main modules. The first, the No-Highlight Detection Module, uses a Variational Autoencoder (VAE) trained only on highlight frames to filter out irrelevant or no-highlight scenes with higher precision than standard autoencoders. The second, the Image Classification Module, is an EfficientNet-based network trained on the newly created Soccer Event (SEV) dataset, which we collected and manually labeled from broadcast matches. The dataset contains 60,000 balanced images across ten soccer event categories. The third, the Fine-Grain Classification Module, employs OSME attention to distinguish between yellow and red cards, which are often visually similar. The SEV dataset is complemented by an independent evaluation set to assess generalization on unseen data. Experiments conducted on 50 full soccer matches from five different leagues, including the UEFA Champions League (UCL), English Premier League, Bundesliga, Brazilian Serie A, and the Iranian Pro League, demonstrate that the proposed system achieves competitive event detection accuracy while maintaining real-time performance. These results suggest that combining open-set filtering with fine-grain attention enhances precision and robustness under realistic broadcast conditions.