Yolov7-Tp: A Deep Convolutional Neural Network for Identifying Tea Buds and Their Plucking Points

To achieve intelligent tea plucking, it is essential to let the machine know the location of tea bud picking. A Yolov7-TP (Yolov7-Tea-bud and plucking point) model is proposed herein for identifying tea buds and their plucking points. Here, images of tea buds in actual tea plantations were collected and labeled to establish a dataset. Thereafter, the Yolov7-TP was trained, tested, optimized, and improved according to the test results. The specific improvements were as follows: a localization loss function of tea-bud plucking points, AW tea plucking-point (AWTP) loss, is proposed to replace the original object key point similarity loss function, which significantly improved the positioning accuracy of plucking points. The shuffle attention mechanism was introduced into the model to advance the detection performance of the model. To evaluate the accuracy of the model to predict tea plucking points, this study proposes a means tea plucking point similarity (MTPS). A new indicator for tea plucking-point evaluation is proposed, the mean tea plucking-point similarity (MTPS), which better reflects the accuracy of the model. Here, the precision, recall, mean Average Precision (Intersection over Union=0.5) (MAP@0.5), and MTPS of the proposed model reached 0.975, 0.884, 0.946, and 0.9835, respectively. The proposed model was observed to be more stable and accurate in a complex environment. This study provides technical solutions and optimization methods for intelligent precision plucking, application values for the acquisition of tea-bud plucking points, and reference values for the identification of key points for other objects.