Adaptive Bidirectional Hybrid Network for Food Image Recognition

In recent years, the rapid development of deep learning has driven its successful application in fields such as food safety and food supervision, attracting substantial attention. Although existing methods have achieved good performance, several challenges remain unresolved: 1) Due to overlapping food items and appearance changes caused by cooking, how to learn discriminative food features needs to be further addressed. 2) Existing methods struggle with limited receptive fields and high redundancy, weakening feature representation. How to expand the receptive field while reducing redundancy remains a challenge. To mitigate the above issues, we develop an adaptive cross-attention network based on convolutional neural networks (CNN) and Vision Transformer (ViT), called ACVCNet. The proposed ACVCNet seamlessly integrates CNN and ViT to leverage their respective strengths. To further enhance feature fusion, ACVCNet incorporates an adaptive weighted cross-attention scheme, dynamically adjusting the contributions to better adapt to diverse food image recognition data. Besides, a dilated attention mechanism is employed in ViT to expand the receptive field and enhance global feature extraction. Similarly, dilated convolution is incorporated into CNN to further enhance feature extraction. Experimental results on four popular food image datasets show that ACVCNet outperforms several recent competitive methods. Specifically, ACVCNet achieves a Top-1 accuracy of 95.544% on VireoFood172, 93.631% on ETHZFood-101, 88.108% on UECFood-100, and 93.101% on UECFOOD-256.