Curvelet-enhanced transformer architecture for blurred action fine-grained detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Accurate fine-grained human behavior recognition is challenging in dynamic videos due to motion blur, occlusion, and illumination variations.
• Existing methods struggle to robustly extract spatial-temporal features under adverse conditions.

Purpose of the Study:

• To propose a novel Multi Curvelet Transformer Network (MCTN) for enhanced human behavior recognition in challenging video scenarios.
• To improve the robustness and accuracy of action recognition models against common image degradations.

Main Methods:

• Developed a motion blur restoration module using the curvelet transform to enhance image clarity.
• Integrated curvelet-based multi-scale attention mechanisms into the Transformer architecture.
• Employed a multi-curvelet transform structure for deeper semantic representation.

Main Results:

• The MCTN achieved a mean average precision (mAP) of 0.822 on benchmark datasets.
• Demonstrated superior performance in fine-grained human behavior recognition compared to existing methods.
• Showcased effectiveness in handling adverse conditions like motion blur and occlusion.

Conclusions:

• The proposed MCTN effectively addresses challenges in human behavior recognition.
• The integration of curvelet transform and Transformer architecture significantly enhances spatial-temporal feature extraction.
• MCTN shows strong potential for real-time intelligent video analysis and human-computer interaction.