Tea Bud Detection and Recognition Based on an Improved YOLOv8 Algorithm

Abstract

In natural tea gardens, tea buds are small, mutually occluded, and surrounded by background texture that closely resembles the target, which causes both missed and false detections. To raise detection accuracy without increasing model complexity, this paper proposes YOLOv8n-HQ, an improved YOLOv8n-based detector for tea buds. The algorithm introduces a P2 high-resolution detection layer to build a four-scale feature pyramid for stronger small-object representation, embeds a convolutional block attention module (CBAM) to suppress background interference, designs a lightweight re-parameterized convolution (LRPC) module to compress the model, and adopts a Wise-IoU loss to improve bounding-box regression accuracy. On a self-built dataset of 15 000 images, the proposed algorithm attains 89.9% mAP@0.5 and 67.6% mAP@0.5:0.95, which are 4.3 and 4.1 percentage points higher than the baseline; the parameter count and computational cost are reduced to 2.05 M and 6.0 GFLOPs, decreases of 35.9% and 31.0%, so that higher accuracy and a smaller model are achieved together. On both detection accuracy and model complexity the algorithm outperforms the compared models.