A lightweight tea buds detection model with occlusion handling

Accurate detection of tea buds plays a crucial role in achieving the automation and intelligence required for tea bud picking. Aiming at the problems that the current object detection algorithms have low detection accuracy for tea buds under occlusion scenarios and the model is too large which is not suitable for mobile deployment, this paper proposed an improved model Yolo-GSG based on Yolox for detecting tea buds. A global context mechanism (GC) was integrated into the feature fusion network to fuse global context information which was used to find tea buds in complex backgrounds. The separated and enhancement attention module (SEAM) was added to the detection head to gather information that was beneficial for tea bud classification while suppressing useless information. The \(\alpha\) $\alpha$ -IoU method was introduced to the post-processing stage to improve missed detection caused by occlusion. The ghost module was introduced to the CSPLayer, considerably reducing the computing and model size. Finally, a tea bud dataset was established for the experiment. The results of the ablation experiment demonstrated that implementing a suitable technique could lead to further enhancements in the model’s performance. The proposed model outperformed models proposed in the previous study (SSD, Faster RCNN, Yolov3, Yolov4, Yolov5, Yolov7) as well as those introduced in subsequent research (Yolov8, RT-DETR, Yolov9), it obtained the best performance with a mAP of 92.71%. The floating-point operations and parameters were reduced to 23.437G and 8.782 M, respectively. In summary, this paper offered both theoretical research and technical support, serving as a foundation for intelligent tea bud picking in real-world scenarios.