Image classification model based on large kernel attention mechanism and relative position self-attention mechanism

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Transformers excel at long-range features but struggle with local details.
• Convolutional Neural Networks (CNNs) capture local details but lack global context.

Purpose of the Study:

• To develop a hybrid model integrating CNN and Transformer strengths for improved image classification.
• To address limitations in capturing both local details and global representations.

Main Methods:

• Proposed a hybrid CNN-Transformer model inspired by Visual Attention Net (VAN) and CoAtNet.
• Incorporated Large Kernel Attention (LKA) into CNN for global feature representation.
• Utilized Transformer blocks with relative position self-attention for local feature detail preservation.

Main Results:

• The hybrid model achieved excellent results on CIFAR-10, CIFAR-100, and Birds400 datasets without additional training data.
• SE_LKACAT model attained 98.01% Top-1 accuracy on CIFAR-10 with only 7.5 million parameters.
• Demonstrated effective combination of local feature accuracy and efficient long-range feature relationship capture.

Conclusions:

• The proposed hybrid model successfully leverages the complementary strengths of CNNs and Transformers.
• Achieved state-of-the-art performance in image classification with a parameter-efficient architecture.