Research on Collar Image Classification Integrating Convolutional Neural Networks and Attention Mechanisms

In this paper, we address the challenges of collar image classification in complex backgrounds, where traditional convolutional neural networks (CNNs) struggle to capture both local details and global contextual information effectively. To overcome these limitations, we propose the MFA-ViT model, which combines the strengths of convolutional operations for local feature extraction with an improved multi-head self-attention mechanism from the Transformer architecture to better model long-range dependencies and global semantic relationships. The proposed MFA module enhances feature diversity and channel attention, while the modified self-attention module reduces computational complexity. We evaluate the performance of the MF A- ViT model on three datasets: Collar6, DeepFashion6, and CIF AR-l O. Experimental results, including ablation studies, demonstrate that our model outperforms traditional CNN s and state-of-the-art methods in both collar-specific classification tasks and general image classification tasks. These results confirm the effectiveness of integrating convolutional structures with Transformer-based attention mechanisms to achieve superior accuracy and efficiency.