A Deep Multi-view Learning method for Rice Grading

摘要

Rice grading has achieved raising attention in recent years for its importance in food security, whereas progress is limited. The difficulty in this topic says that the rice kernels are crowed in the visual field of, say a camera, which makes the detection of a single kernel hard. In this paper, we are based on a newly designed rice streaming system and propose a novel rice grading model. The streaming system snapshots a kernel from three different visual directions, producing three images of a single kernel. The FIST-Model analyses these images by employing a multi-view learning method which minimizes the information loss and generates an intact representation of the rice kernel. Such a representation remains strong discriminability and is effective to determine the grading level of the rice. Finally, we evaluate the performance of the proposed FIST-Model on the FIST-Rice dataset by setting a series of experiments based on different deep learning models. The result indicates that the FIST-Model has a superior performance over previous rice grading model.