Extracting tea bud contour and location of picking points in large scene using case segmentation

Automatic picking tea buds is ever increasing in the continuous expansion of famous brand tea market at present. However, manual picking cannot fully meet the needs of short-term concentrated harvesting in the large-scale tea production. Mechanical mass picking can also share the very short window period of famous brand tea. Therefore, it is the urgent demand to realize the intelligent and accurate picking of famous tea. Among them, visual identification has been confined to the small target of tea buds and the complex background under the natural environment in the field. Fortunately, Yolov5s-segment network model is suitable to identify tea buds, due to the easy deployment, simple integrated structure and real-time detection. In this study, a novel model was proposed to extract the contour and then locate the picking point of tea bud using improved Yolov5s-segment. The rapid recognition and accurate location of picking points were realized for the tea buds in large field scenes. Firstly, P2 micro-target detection layer was imported into the Neck and Head of the original network. The small targets detection of P3, P4 and P5 layers was achieved in the improved Yolov5s-segment network. Secondly, the contour features of tea bud were extracted in the natural lighting environment. CBAM (convolutional block attention module) was added into the end of Backbone of Yolov5s-segment network for the anti-interference of the improved model. Finally, the position coordinates of the bud stem were extracted, according to the bud contour. The picking points were located accurately for the single bud, one bud with one leaf, and one bud with two leaves. Furthermore, the micro target detection layer and CBAM attention mechanism were gradually added into the original Yolov5s-segment model. In ablation test, the CBAM attention mechanism was replaced with SE, ECA and SimAM. After that, the optimal model was compared with Yolov8s-segment and Yolact. Finally, the positioning test was carried out using the optimal model. The results showed that the best performance was achieved in the model with the micro-target detection layer and CBAM attention mechanism. Compared with the original Yolov5s-segment model, the accuracy, recall rate, F1 score, mean average precision mAP₅₀ and mAP_50-95 were improved by 7.0, 8.9, 8.1, 8.3 and 7.3 percentage points, respectively. In the comparison test, the mAP₅₀ of the improved model increased by 4.2 and 9.5 percentage points, respectively, compared with the Yolov8s-segment and Yolact. The positioning test showed that the pixel coordinates of picking points were accurately obtained in the 0° and 45° shooting angles of camera under different states, indicating the high positioning accuracy. The improved model can be used to accurately extract the outline of single bud, one bud with one leaf and one bud with two leaves in large field scenes. The precise location of picking points can be realized at the same time. The findings can provide a theoretical basis for intelligent and rapid picking of famous tea.