OM-VST: A video action recognition model based on optimized downsampling module combined with multi-scale feature...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current video classification models struggle with accuracy due to complex video data and high parameter counts.
• Challenges include subtle category differences, background noise, and illumination variations, leading to suboptimal performance.
• High parameter counts result in extended training times and increased energy consumption.

Purpose of the Study:

• To introduce an improved video classification model, the OM-Video Swin Transformer (OM-VST).
• To enhance feature perception and characterization capabilities for more accurate video analysis.
• To reduce model complexity and computational overhead in video classification tasks.

Main Methods:

• Developed the OM-Video Swin Transformer (OM-VST) by integrating a multi-scale feature fusion module and an optimized downsampling module.
• Built upon the existing Video Swin Transformer (VST) architecture.
• Conducted comparative experiments against mainstream models like VST, SlowFast, and TSM on a public dataset.

Main Results:

• The OM-VST model achieved a 2.81% improvement in classification accuracy compared to existing models.
• Significantly reduced the number of model parameters by 54.7%.
• Demonstrated superior performance in accurately identifying and labeling video content.

Conclusions:

• The OM-VST model effectively addresses the limitations of current video classification approaches.
• Achieves higher accuracy and greater efficiency, making it suitable for practical applications.
• Offers a promising advancement in automated video content analysis and understanding.