To enhance the intelligence level of mechanical tea harvesting, the author designed a tea harvesting experimental platform consisting of a support frame, arc-shaped harvesting blade, blade screw lifting plate, 4 rollers, 2 drive motors, controller, and battery pack. Using YOLOv5s 6.0 as the baseline model, several modifications were implemented: the backbone network was replaced with MobilenetV3; a CBAM attention module was integrated before the detection layer; the lightweight universal upsampling operator CARAFE was adopted to substitute the nearest neighbor interpolation method. Furthermore, by incorporating trade-off functions and enhancing the CIOU loss function, a novel mathematical model YOLOv5s+ was developed for tea leaf detection. Subsequently, tea bud images taken at different heights(10, 20, 30, 40, 50 cm) and angles(15°, 30°, 45°, 60°, 75°, 90°) were used as samples to test their impact on network recognition accuracy. The results demonstrated that optimal model performance was achieved when images were acquired at a 20 cm vertical distance from the tea tree canopy with a 45° shooting angle. Using the image set captured under these parameters for ablation experiments, YOLOv5s+ achieved mean average precision and recall rates of 0.935 and 0.912 respectively for tea bud recognition, showing improvements of 2.97% and 2.82% compared to YOLOv5s.