Detection of Foreign Objects overlapped to Green Onion Flakes

Abstract

Foreign objects in food can cause disgust in consumers as well as have a direct impact on health. With the recent development of image recognition technology using deep learning, many studies are being conducted to detect foreign objects in food through deep learning. Deep learning can learn features well in roughly uniform distributions of class labels. However, the classes of foreign objects are diverse and difficult to collect industrial site. As a result, there is a problem with the distribution of long-tailed data with a large number of normal classes and a few abnormal classes. Moreover, even though deep learning, adjacent objects are difficult to classify because their boundaries are ambiguous. In this study, we focus on finding foreign objects overlapped to the green onion flakes that are the base material used in many countries. To detect foreign objects (e.g. insect, hair, etc.) overlapped to green onion flakes, we develop artificial minority over-sampling method. Through this method, training data is generated for foreign objects overlapped to green onion flakes. Our network classified images of foreign objects overlapped to green onion flakes 94.29% success ratio among a total of 105 objects. The results show that when trained with the proposed re-sampling, the network is able to achieve significant performance gains on foreign objects overlapped to green onion flakes.