Food Image Segmentation Based on Deep and Shallow Dual-branch Network

Food image segmentation is an important research area within the fields of computer vision and machine learning. Traditional methods often input high-resolution food images at large sizes directly into network models, which leads to high computational costs. Additionally, effectively distinguishing between different foods with similar appearances and the same food in different forms poses a significant challenge. This paper introduces a dual-branch structure network based on Swin Transformer and convolutional neural networks (FDSNet), which significantly reduces the computational costs of processing large-size input images. Furthermore, this study introduces a multi-scale feature fusion technique that effectively integrates feature information from different scales and levels, enabling the model to more accurately segment and recognize different foods. Our method can more precisely perform food image segmentation, helping people improve their diets and manage health better. Training and testing on the FoodSeg103 and UECFoodPixComplete public food datasets have shown that our model achieves mean Intersection over Union (IoU) scores of 47.34 and 75.89, respectively, demonstrating higher accuracy and computational efficiency compared to other methods. Our code is released at https://github.com/llevelingup/FDSNet.