Deep Learning-Based Object Detection of Barley Seeds

Abstract

Computer vision techniques have been widely used in the food manufacturing industry today due to their speed, convenience, and low labor cost. As an important raw material in beer, barley seeds largely determine the flavor and taste of the beer brewing process. To ensure the quality of beer, beer brewers will strictly screen the varieties of barley seeds and ensure the purity of malt. Traditional manual detection needs a lot of professional training, and because of the high similarity between different kinds of barley seeds, tedious and time-consuming manual detection leads to a high error rate. However, chemical testing requires professional equipment, reagents, and laboratories, which have high-cost performance. Thus, an efficient and accurate object detection technique is considered to replace manually distinguishing barley seed types. This thesis uses the deep learning-based object detection network YOLOv3 model to help automatically and accurately detecting barley locations and identifying seed types from iPhone-based images of barley seeds. The barley seed samples used in this project are all provided by our industrial collaborator, and captured by iPhone 11 or iPhone 11 pro in high-definition resolutions. A total of nine varieties of barleys are trained in this study, and the data set includes images of a single grain and images of multiple grains (multi-categories). In this experiment, the best mAP (mean Average Precision) value we obtained is 97.1%, the model recognition (localization and classification) precision is 91.2%, and the recall rate is 95.9%.