DKTransformer: An Accurate and Efficient Model for Fine-Grained Food Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Food image classification is crucial for dietary analysis and health computing.
• Fine-grained food categories present challenges due to visual similarity and intra-class variation.
• Existing methods struggle with complex food scenes by over-relying on global or local features.

Purpose of the Study:

• To propose DKTransformer, a lightweight hybrid architecture for fine-grained food image classification.
• To enhance local detail modeling and capture long-range dependencies efficiently.
• To reduce computational cost and parameter redundancy in food image classification models.

Main Methods:

• Developed DKTransformer, integrating Vision Transformers (ViT) and Convolutional Neural Networks (CNNs).
• Introduced a Local Detail Extraction (LDE) module using depthwise separable convolution.
• Designed a Multi-Scale Dilated Attention (MSDA) module for efficient long-range dependency modeling.
• Employed an Efficient Kolmogorov-Arnold Network (EfficientKAN) to replace conventional feedforward networks.

Main Results:

• DKTransformer achieved 92.71% Top-1 accuracy on ETH Food-101 with 47M parameters.
• Attained 90.70% Top-1 accuracy on Vireo-Food-172 and 66.89% on ISIA Food-500.
• Demonstrated strong generalization across diverse food datasets and scales.
• Showcased a favorable balance between classification accuracy and computational efficiency.

Conclusions:

• DKTransformer effectively addresses the challenges of fine-grained food image classification.
• The hybrid architecture offers a promising solution for accurate and efficient food recognition.
• The method shows significant potential for applications in dietary analysis and health computing.