A lightweight edge-deployable detection method for sugarcane stem nodes in intelligent seed cane cutting equipment

Accurate detection of sugarcane stem nodes is essential for intelligent seed cane cutting, yet existing detectors are often limited by large parameter counts, small target feature loss, and performance degradation under motion blur, making them difficult to deploy on edge devices.

This study developed SAD-YOLOv11, a lightweight and accurate detector based on YOLOv11, incorporating three modifications. These include an average pooling downsampling module named ADown for reducing computational cost while preserving edge features, a self attention based AIFI module for global contextual modeling, and a SEAMHead detection head with joint channel and spatial attention for adaptive feature calibration. The model was evaluated on a custom sugarcane stem node dataset.

The model achieved a precision of 95.92%, a recall of 95.66%, and an mAP50 of 97.57%, maintaining detection accuracy at a level comparable to the baseline YOLOv11 while reducing parameter count and FLOPs by 18.13% and 17.55%. Under physically captured motion blur conditions, the model maintained more stable detection than the baseline, with the advantage becoming more pronounced as blur severity increased. Deployment on the Jetson Orin NX platform with TensorRT-FP16 quantization achieved a throughput of 37.1 FPS with a zero frame drop rate at a 30 FPS input frame rate, and compressed the engine size by 46.7%.

These results suggest that SAD-YOLOv11 achieves a sound balance between detection accuracy and computational efficiency suitable for intelligent seed cane cutting equipment.