Scholarlink

<p>To address the limitations of traditional Generative Adversarial Networks (GANs) in boxing video action prediction, such as insufficient multi-scale feature extraction, insufficient action diversity, and difficulty in distinguishing similar actions, an enhanced model incorporating three key techniques is proposed. The model employs multi-scale fusion to optimize both the generative and discriminative networks: The generator uses a U-shaped structure with multi-scale convolution blocks to capture global and local features, while the discriminator utilizes pyramid pooling and 3D convolution for multi-scale analysis. Additionally, a Convolutional Block Attention Module (CBAM) is introduced to emphasize key action features, and contrastive learning is integrated to enhance feature discrimination via positive and negative sample pairs. The ablation experiment confirmed the contribution of each component, and the comparative experiment proved the superiority of the model. On the Boxing-UCF101 dataset, it achieved 90.3% accuracy, 41.2&#xa0;dB PSNR, and 10.5 SD. On the more challenging Kinetics-Boxing dataset, it attained 84.2% accuracy, 39.5&#xa0;dB PSNR, and 11.3 SD. This model had only 13.8&#xa0;M parameters and 52.1&#xa0;ms inference time per frame, providing an efficient and practical solution for boxing action prediction.</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

Boxing video action prediction model based on contrastive learning and multi-scale fusion

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Qingdong Lin,
  • Yuliang Li,
  • Jingqi Sun

摘要

To address the limitations of traditional Generative Adversarial Networks (GANs) in boxing video action prediction, such as insufficient multi-scale feature extraction, insufficient action diversity, and difficulty in distinguishing similar actions, an enhanced model incorporating three key techniques is proposed. The model employs multi-scale fusion to optimize both the generative and discriminative networks: The generator uses a U-shaped structure with multi-scale convolution blocks to capture global and local features, while the discriminator utilizes pyramid pooling and 3D convolution for multi-scale analysis. Additionally, a Convolutional Block Attention Module (CBAM) is introduced to emphasize key action features, and contrastive learning is integrated to enhance feature discrimination via positive and negative sample pairs. The ablation experiment confirmed the contribution of each component, and the comparative experiment proved the superiority of the model. On the Boxing-UCF101 dataset, it achieved 90.3% accuracy, 41.2 dB PSNR, and 10.5 SD. On the more challenging Kinetics-Boxing dataset, it attained 84.2% accuracy, 39.5 dB PSNR, and 11.3 SD. This model had only 13.8 M parameters and 52.1 ms inference time per frame, providing an efficient and practical solution for boxing action prediction.