Design and operation of a deep-learning-based fresh tea-leaf sorting robot

Abstract

Tea is one of the most popular beverages worldwide. It is rich in substances, such as tea polyphenols and polysaccharides, that are closely related to human health. However, fresh tea leaves picked by machines are neither neat nor uniform. In this study, we apply a deep learning algorithm and the DELTA parallel robotic arm for high-precision and fast sorting of machine-picked fresh tea leaves. First, a convolutional neural network algorithm and a regional segmentation method are used to achieve fast identification and localization of machine-picked fresh tea leaves. Second, the actual running time model of the stepper motor at different pulse frequencies was established by the interpolation fitting method. The running times of the parallel mechanical arm associated with S-curve type acceleration and deceleration motions were calculated accurately using this model, and the sorting point of tea fresh leaves was then determined. A multi-objective, continuous sorting model of machine-picked fresh tea leaves is constructed to verify and optimize the model effects. The results of the robot’s arm running-time test show that the running time from the end of the robotic arm’s application to the fresh tea leaf sorting plane is in the range 500–1800 ms. The experimental results of fresh tea leaf recognition and classification show that after 150 iterations, the recognition accuracy of the validation set can reach 99.82%. Finally, the average sorting accuracy of the four experiments reach 89%, with the highest sorting accuracy reaching up to 92%. These results demonstrate that the proposed method leads to an excellent sorting effect pertaining to machine-picked fresh tea leaves. This approach could be easily applied to any agricultural product sorting application.