DMSAL-Net: A Two-Stage Network with Deblurring, Multi-Scale Feature Enhancement, and Morphologically Adaptive Localization for Tea Shoot Picking Points

To achieve high-precision and real-time automated tea shoot picking under complex natural field scenarios, this study proposes a two-stage tea shoot picking point localization network, DMSAL-Net, to address inherent challenges including image defocus blur, arbitrary target scale variations, cluttered background interference, and diverse tea shoot morphological deformations. Unlike existing tea detection and keypoint localization methods that suffer from poor anti-interference capability and limited adaptability to morphological changes, the proposed framework presents three major methodological and scientific contributions. First, an enhanced detection network, MSC-Net, is constructed based on an improved YOLOv5 architecture, in which the MRT-Net sharpness restoration module recovers blurred field images, the embedded multi-scale dilated attention (MSDA) module enhances multi-scale feature perception, and spatial–channel reconstruction convolution (SCConv) suppresses complex background redundancy, collectively improving detection robustness in unconstrained environments. Second, an improved YOLOv11-Pose-based AdaPoint-Net is developed by incorporating linear deformable convolution (LDConv) to effectively model variable tea stem bending and morphological differences, thereby overcoming the fixed-convolution limitations of conventional keypoint localization methods. Third, channel pruning is implemented for lightweight optimization to balance localization accuracy and real-time inference performance. Quantitative experiments demonstrate that the proposed DMSAL-Net achieves a detection precision of 94.6%, a recall of 93.6%, and an mAP50 of 97.1%. Ten-fold cross-validation yields a [email protected] of 87.03% ± 1.06%, representing an average localization accuracy improvement of 18.52 percentage points over the YOLOv11-Pose baseline. The proposed method outperformed the YOLOv11-Pose baseline in all ten held-out test folds, demonstrating a consistent performance advantage across the predefined data partitions. Moreover, the designed filtering strategy improves [email protected] by 3.67 percentage points. After channel pruning, the model size and parameter count are reduced by 49.6% and 76.0%, respectively, while the inference speed increases from 4.76 FPS to 16.72 FPS with only a marginal accuracy decrease. Cross-cultivar and cross-season generalization experiments further demonstrate strong robustness and generalization capability, confirming that the proposed DMSAL-Net provides an effective technical solution for accurate and efficient automated tea shoot harvesting in complex natural environments.