Deep-Rice: Deep Multi-sensor Image Recognition for Grading Rice*

Rice grading is an important topic in the research of food security, which targets at assessing the quality of rice. However, there is seldom attention from researchers. In this paper, we propose a novel rice grading system, named as Deep-Rice, which is built upon a deep learning framework. Specifically, Deep-Rice employs a multi-view CNN architecture to extract the discriminative features from different views of rice images, and tries to optimize the CNN parameters by using a modified softmax loss function. In accompany with this deep model, we also build a large-scaled rice dataset, which is denoted as FIST-Rice, to provide a baseline database for the research of food security. FIST-Rice is the first publicly available multi-sensor rice dataset, which contains sound and unsound kernel samples with the total number of 30,000. Each sample is captured under three different illumination conditions. We evaluate the proposed Deep-Rice model on the FIST-Rice dataset by comparing with the AdaBoost and SVM methods. The experimental results indicate the Deep-Rice model achieves improved performance in different conditions of light intensity. The FIST-Rice dataset and baseline codes will be released for further public research.