Advancing lightweight and efficient detection of tomato main stems for edge device deployment

Automated pruning and defoliation of tomato plants are essential in modern cultivation systems for optimizing canopy structure, enhancing air circulation, and increasing yield. However, detecting main stems in the field faces significant challenges, like complex background interference, limited field of view, dense foliage occlusion, and curved stems. To address those challenges while ensuring hardware friendliness, computational efficiency, and real-time response, this study proposes a lightweight tomato main stem detection, optimisation, and deployment scheme. First, an efficient semi-automatic rotated bounding box annotation strategy is employed to segment the visible main stem segments, thus improving the adaptability to curved stems. Second, the lightweight network, YOLOR-Slim, is constructed to significantly reduce model complexity while maintaining detection performance through automated iterative pruning at the group-level of channel importance and a hybrid feature-based and logic-based knowledge distillation mechanism. Finally, an efficient and real-time main stem detection is achieved by deploying the model on inference engines and embedded platforms with various types and quantization bits. Experimental results showed that YOLOR-Slim achieved 87.5 % mAP@50, 1.9G Flops, 1.4 M parameters, and 7.4 ms inference time (pre-processing, inference, and post-processing) on the workstation, representing reductions of 2.8 %, 10.0 M, and 27.5G compared to the original model. After conversion with TensorRT, the inference time on Jetson Nano reached 57.6 ms, validating the operational efficiency and deployment applicability on edge devices. The YOLOR-Slim strikes a balance between inference speed, computational resources usage, and detection accuracy, providing a reliable perceptual foundation for automated pruning tasks in precision agriculture. • Lightweight tomato main stem detection application for resource-constrained edge devices. • Semi-automated and segmented rotated bounding box labelling to represent main stems. • Automated network pruning and knowledge distillation for light-weighting and efficiency. • Inference engine deployment for resource constrained embedded platforms.