Improved deep learning image classification algorithm based on Swin Transformer V2.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel at extracting local image features but struggle with global dependencies due to limited receptive fields.
• Transformers effectively model global dependencies but lack inherent mechanisms for local information exchange within specific regions.
• Existing models face challenges in simultaneously optimizing local feature extraction and global dependency modeling for image classification.

Purpose of the Study:

• To enhance the Swin Transformer V2 model by synergistically combining the strengths of CNNs and Transformers.
• To improve the model's ability to capture both local image features and long-range dependencies.
• To boost image classification accuracy and generalization capabilities.

Main Methods:

• Integration of convolutional operations and self-attention mechanisms into the Swin Transformer V2 architecture.
• Introduction of Swin Transformer Stem, inverted residual feed-forward network, and Dual-Branch Downsampling for enhanced local information extraction.
• Application of downsampling to the attention mechanism's Query (Q) and Key (K) to reduce computational and memory overhead.

Main Results:

• Significant improvements in classification accuracy across multiple image classification datasets under identical training conditions.
• Demonstrated enhanced extraction of local information through the novel architectural components.
• Showcased more robust generalization capabilities compared to baseline models.

Conclusions:

• The proposed hybrid approach effectively leverages the complementary strengths of CNNs and Transformers for superior image classification.
• The enhanced Swin Transformer V2 model offers a more comprehensive feature representation by integrating local and global modeling.
• The method provides a promising direction for developing more powerful and efficient deep learning models for computer vision tasks.