Meta-Learning-Based Lightweight Method for Food Calorie Estimation

Abstract

As a significant research component in nutritional assessment, vision-based food calorie estimation has been studied and applied due to its higher accuracy and efficiency. In this paper, a lightweight network for food calorie estimation is designed, called MeLL-cal. Firstly, a feature extraction module is proposed based on meta-learning ideas to generate informative representations, such as color, texture, and edge features, for unseen foods. Secondly, within the feature extraction module, a large convolutional kernel is proposed to provide a larger receptive field, which aims to capture more shape and semantic information and minimize information loss. Then, to achieve efficient calorie estimation with lower computational complexity, the calorie estimation module employs query-based inference to achieve optimal feature expression. Additionally, an adaptive fine-tuning module is also designed to refine estimation accuracy according to different datasets. The extensive experiments demonstrate the superiority of the MeLL-cal in terms of a PMAE of 18.7% and 31.1%, respectively, with only 2.313K parameters and 1.036 ms inference time on the Menu match dataset and the Calo world dataset.