Adapting New Categories for Food Recognition with Deep Representation

Learning to classify new (target) data in a different domain is always an interesting and challenging task in data mining. The classifier could suffer the dataset bias when predicting the new categories from target domain. Many adaptation methods have been proposed to adjust this bias but are limited to using data either from similar categories or requiring a large number of labeled examples from the target domain. Automatically adapting and recognizing new food categories is a very practical task in daily life. In this paper, we propose a new method that can alleviate the dataset bias for food image recognition. To obtain less biased feature representation from the food images, we fine-tuned GoogLeNet as our deep feature extractor and achieve state-of-the-art performance on the Food-101 dataset. Using the deep representation, our method can learn efficient classifiers with fewer labeled examples. More specifically, our method employs an external classifier for adaptation, called "negative classifier". Experiment results show that utilizing the parameters of the negative classifier, our method can achieve better performance and converge faster to adapt the new categories.