A Comparative Survey of Vision Transformers for Feature Extraction in Texture Analysis

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Texture is a key visual feature in image analysis and pattern recognition.
• Convolutional Neural Networks (CNNs) are established methods for texture analysis.
• Vision Transformers (ViTs) show promise in broader visual recognition tasks.

Purpose of the Study:

• Investigate the effectiveness of Vision Transformers (ViTs) for texture recognition.
• Analyze the capabilities and limitations of various ViT architectures as feature extractors.
• Compare ViT performance against CNN-based and hand-engineered approaches.

Main Methods:

• Evaluated 25 ViT variants as feature extractors for texture analysis.
• Compared accuracy and computational efficiency of ViTs, CNNs (ResNet50), and hand-engineered methods.
• Utilized in-the-wild texture datasets and strong pre-training strategies.

Main Results:

• ViTs generally outperformed CNNs and hand-engineered models in texture recognition accuracy.
• BeiTv2-B/16 achieved the highest average accuracy (85.7%), followed by ViT-B/16-DINO (84.1%) and Swin-B (80.8%).
• Despite higher FLOPs/parameters, ViTs like ViT-B and BeiT(v2) offered faster GPU feature extraction than ResNet50.

Conclusions:

• ViTs are a powerful and efficient tool for texture analysis, surpassing traditional methods.
• Strong pre-training enhances ViT performance, especially on diverse, real-world texture datasets.
• Future research should focus on ViT efficiency improvements and domain-specific adaptations for texture recognition.