<i>Application of Deep Learning Algorithm on Tea Shoot Identification and Localization </i>

Abstract

Abstract. Tea (Camellia sinensis) is one of the most widely consumed beverage in Taiwan. More than 40 thousand tons of tea is consumed per year. However, the amount of tea production declined dramatically in recent years due to the serious labor shortage situation. Although mechanical tea harvester boosts the harvesting efficiency by 12 to 15 times than hand-picking method, it cannot avoid broken and old leaves. High-quality tea leaves are mostly made from full leaves, or even plucked with specific composition, such as one bud with one or two leaves, which is defined as tea shoot. Only hand-picking operation can meet the criteria so that its harvesting is highly relied on the labor availability. Automatic identification and localization of tea plucking point not only help alleviate the situation of labor shortage but improve the quality of mechanical tea harvesting. In this study, the primary purpose is to develop an algorithm to localize the plucking points of tea leaves. The algorithm consists of two parts: identifying specific morphology of tea shoot and localizing the position for plucking. Deep learning algorithm, Faster R-CNN, one of the main stream region proposal methods, is applied to find the features of tea leaves, and a pre-trained ZF model is adopted. The average precision (AP) of the identification model can achieve 86.04%. Preliminary results demonstrate successful identification and localization of the specified tea shoots for two varieties of tea cultivars (Taiwan Tea Experiment Station, TTES No.8 and TTES No.18).