AGTH-Net: Attention-Based Graph Convolution-Guided Third-Order Hourglass Network for Sports Video Classification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sports video classification is vital for applications like smart TV and video on demand.
• Current manual classification methods are prone to human error and inaccuracy.
• There is a need for automated, accurate sports video classification techniques.

Purpose of the Study:

• To propose a novel neural network model for automatic sports video classification.
• To address the limitations of manual classification by introducing an attention-based graph convolution approach.
• To enhance the accuracy and efficiency of sports video analysis.

Main Methods:

• Developed an attention-based graph convolution model for neighborhood node weight allocation, minimizing error node impact.
• Utilized a third-order hourglass network structure for multiscale feature extraction and fusion in complex sports videos.
• Incorporated residual-intensive modules within the hourglass network for improved feature transfer and reuse.

Main Results:

• The proposed AGTH-Net model demonstrated superior performance in sports video classification tasks.
• Ablation experiments confirmed the effectiveness of the attention mechanism and hourglass structure.
• The model successfully extracts and fuses multiscale features, enhancing classification accuracy.

Conclusions:

• The AGTH-Net model offers a significant advancement in automatic sports video classification.
• The integration of attention mechanisms and hourglass networks provides a robust solution for complex video data.
• This research paves the way for more reliable and efficient sports video analysis systems.