ChineseFoodNet: A large-scale Image Dataset for Chinese Food Recognition

In this paper, we firstly introduces a new and challenging large-scale food image dataset called "ChineseFoodNet", which aims to automatically recognizing pictured Chinese dishes. Importantly, unlike the existing datasets, in which images are either from recipe pictures or selfie, images of each category consists of not only from recipe pictures but also real dishes. Furthermore, we present our effort in building the ChineseFoodNet, that covers not only data collection such as different views and backgrounds and rich attitudes, but also building the ChineseFoodNet combining with machine learning and crowdsourcing methods . Secondly, we propose a novel deep convolutional neural network~(CNN) referred as to "TastyNet" for Chinese dish recognition. Our algorithm is benchmarked on the ChineseFoodNet and achieves top-1 accuracy of 72.55\% and top-5 accuracy of 91.31\%, respectively.The latest dataset can be downloaded at \url{https://sites.google.com/view/chinesefoodnet/}.