Automated food recognition has great value in terms of health benefits, as it provides a more efficient and accurate method of recording an individual's diet. It is the high variability within the same food groups and yet, the subtle distinctions between different food groups that makes food recognition a difficult task.
In this work, we evaluated and compared three popular image descriptors for their performance on two challenging food data sets: 50 Chinese Foods and ImageNet Foods, the latter of which we collected from ImageNet. We also proposed a method to automatically discover food related attributes to give rise to a recipe-based food descriptor. Our food descriptors were able to improve the performance of the baseline descriptors by encoding attributes such as ingredients, cooking method, and nutritional content. We also demonstrated that our food descriptors are capable of knowledge transfer for zero-shot learning.
Keywords: image recognition, classification, fine-grained recognition