STEFF: Spatio-temporal EfficientNet for dynamic texture classification in outdoor scenes.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Dynamic texture classification is crucial for computer vision but challenging due to complex spatio-temporal dynamics.
• Existing methods struggle with the inherent variability of dynamic textures in real-world scenes.

Purpose of the Study:

• To propose a novel spatio-temporal approach (STEFF) for robust dynamic texture classification.
• To integrate appearance and motion features using deep learning for improved classification accuracy.

Main Methods:

• Developed a spatio-temporal approach (STEFF) combining difference and average operators on video sequences.
• Extracted deep texture features from outdoor scenes and integrated them into an EfficientNet model.
• Employed fine-tuning and regularization techniques for model optimization.

Main Results:

• Achieved high classification accuracies: 95.95% on Yupenn, 94.09% on DynTex++, and 98.01% on Yupenn++.
• Demonstrated superior performance compared to existing dynamic texture classification models.
• Validated the approach's effectiveness and efficiency across multiple datasets.

Conclusions:

• The proposed STEFF method offers a robust and effective solution for dynamic texture classification in computer vision.
• Integrating spatial and temporal features via deep learning significantly enhances classification performance.
• The approach shows strong potential for real-world applications involving dynamic scene analysis.