A transformer-based path fusion detector network for tea bud detection

Timely and accurate detection of tender tea buds is essential for premium tea production, yet remains a highly labor-intensive task in modern plantations. With China’s mechanization rate below 10% and rising labor costs driven by an aging workforce, intelligent robotic harvesting has become imperative. Reliable tea bud detection is the critical perception module that guides robotic arms to precise picking locations. However, natural tea gardens present severe challenges, including small target size, dense clustering, frequent occlusions by overlapping leaves, high color and texture similarity to background foliage, and varying illumination conditions. To address these issues, we propose TPFD-Net, a novel Transformer-based Path Fusion Detector Network that integrates multi-scale feature fusion, global context modeling, and anchor-free regression. Specifically, the Attentive Depthwise Fusion Cascade (ADFC) module aggregates rich semantic and spatial information across scales, a Transformer encoder captures long-range dependencies for robust recognition under complex backgrounds, and lightweight anchor-free heads enable efficient multi-scale detection. We also construct and publicly release a high-quality tea bud dataset comprising 3,963 high-resolution images captured under diverse illumination, cultivars, and growth stages. Experiments on a self-constructed JinQiu tea dataset and two public open-source datasets demonstrate that TPFD-Net achieves improved recall and competitive mAP@0.5 under both primary benchmark and external validation settings while maintaining lower inference latency than state-of-the-art YOLO-based detectors. On the NVIDIA Jetson Orin NX edge platform, TPFD-Net achieves 28.73 FPS at an input resolution of 320 × 320, corresponding to approximately 34.8 ms per image. These results verify the near real-time inference capability of TPFD-Net and demonstrate its potential for robotic tea harvesting systems.