Food Image Classification Based on Residual Network

Abstract

As food culture and internet technology evolve, tracking the nutritional information of daily food intake becomes increasingly important for assessing dietary habits and health management status. However, effective food image classification is a prerequisite. Food images present a fine-grained image recognition problem characterized by large inter-class differences and small intra-class differences. The presence of mutual occlusion between foods and background noise challenges existing food image classification techniques in extracting robust visual features. In response to these challenges, this paper proposes a food image classification residual network incorporating pyramid segmentation attention and soft thresholding. Attention is applied across both spatial and channel dimensions to mitigate the impact of noisy data on classification results. In each improved residual block, pyramid segmentation attention (PSA) is embedded to replace the convolutional unit, extracting target features through the spatial-level visual attention vector and multi-scale response map. Concurrently, a soft thresholding sub-network is embedded within the basic module of the network, employing channel attention to automatically learn the threshold for each sample, thereby suppressing redundant information in the image. Multiple experiments were conducted using the VireoFood-251 food dataset, with results indicating a classification accuracy of 87.03%. When compared to classical models ResNet34 and ResNet50, the accuracy increased by 5.71% and 3.02%, respectively, validating the feasibility of the proposed network framework.