EdgeSugarcane: a lightweight high-precision method for real-time sugarcane node detection in edge computing environments.

Accurate detection of sugarcane nodes in natural environments is crucial for realizing intelligent sugarcane cutting and precise planting localization. However, current sugarcane node detection models often face challenges such as large parameter sizes, poor adaptability to deployment environments, and limited real-world detection accuracy. To address these challenges, this research proposes a high-precision and lightweight EdgeSugarcane detection model. Firstly, based on YOLOv8, an improved EdgeSugarcane model is proposed. By introducing an interactive attention mechanism across channel and spatial dimensions, the model's ability to represent node features is enhanced. Then, combined with TensorRT acceleration and optimization, the optimal FP16 quantization deployment scheme is proposed. Finally, end-to-end deployment is implemented on the NVIDIA Orin NX edge device, and its performance and resource consumption in practical applications are analyzed in depth. The experimental results show that EdgeSugarcane achieves a precision of 0.935, a recall of 0.8, and a mAP of 0.87 on the test set, with a model size of 89.9 MB. Compared to YOLOv8, the mAP is improved by 0.6%, and the inference speed is increased by 44%. With lossless precision, the inference time after FP16 quantization is only 1.9ms, a 3.3-fold improvement compared to before optimization, and the model size changes very little. On the NVIDIA Orin NX device, the single-frame inference, pre-processing, and post-processing times are 1.5ms, 60.6ms, and 4.4ms, respectively. The EdgeSugarcane model demonstrates excellent real-time performance and high accuracy under natural field conditions, offering a viable solution for integration into edge-based robotic systems for intelligent sugarcane cutting and precision planting.