DBRS-Net: A Hybrid Framework for Food Image Classification

Abstract

Food image classification holds significant importance in applications such as food retrieval, nutritional assessment, and dietary management. However, food images typically exhibit pronounced intra-class variations, high inter-class similarity, and complex shooting environments, limiting the classification performance of traditional deep learning approaches. To address this, this paper proposes a novel hybrid network architecture, the Dual-Branch ResNet-Swin Network (DBRS-Net). By integrating ResNet with the window-based attention mechanism of Swin Transformer, it leverages the complementary strengths of both architectures in local fine-grained feature extraction and global context modelling. Specifically, the ResNet branch captures local features such as texture and shape, while the Swin Transformer branch learns overall structure and long-range dependencies. A simple yet effective feature fusion strategy then synthesises comprehensive image representations. Experimental results demonstrate that even without introducing additional complex modules, DBRS-Net achieves stable classification performance on the Food11 dataset, attaining a top-1 accuracy of 94.83%. This provides a reliable foundation for research into food image recognition on larger-scale or more diverse datasets.