Abstract & Overview
Generalization of food image recognition frameworks is difficult due to the wide variety of food categories in cuisines across cultures. The performance of the deep neural network models highly depends on the training dataset. To overcome this problem, we propose to extract context information from images in order to increase the discrimination capacity of networks. In this work, we utilize the CLIP architecture with the automatically derived ingredient context from food images. A list of ingredients are associated with each food category, which is later modeled as text after a voting process and fed to a CLIP architecture together with input image. Experimental results on the Food101 dataset show that this approach significantly improves the model’s performance, achieving a 2% overall increase in accuracy. This improvement varies across food classes, with increases ranging from 0.5% to as much as 22%. The proposed framework, CLIP fed with ingredient text, outperforms Yolov8 (81.46%) with 81.80% top 1 overall accuracy over 101 classes.