Improved YOLOv11n-seg for Impurity Detection in Mechanically Harvested Sugarcane

Fengguang He^1,2Sili Zhou^1,2Pinlan Chen^1,2*Ganran Deng^1,2Shaobo Feng^3*Guojie Li^1,2Zhende Cui^1,2Shuang Zheng^1,2Ling Li^1,2Bin Yan^1,2Shuangmei Qin^1,2Xilin Wang^1,2Ye Dai^1,2Zehua Liu^1,2

• ¹Agricultural Machinery Research Institute, Chinese Academy of Tropical Agricultural Sciences, Zhanjiang, China
• ²Key Laboratory of Tropical Agricultural Machinery, Ministry of Agriculture and Rural Affairs, Zhanjiang, China
• ³Guangxi Research Institute of Metrology & Test, Nanning, China

The content of impurities in mechanically harvested sugarcane is crucial for assessing harvest quality and sale price. To enable intelligent detection of impurities in mechanically harvested sugarcane, this study proposes a sugarcane impurity detection method based on an Improved YOLOv11n-seg model. The method incorporates four enhancement modules into the YOLOv11n-seg architecture. Firstly, a lightweight C2_Ghost module is introduced in the high-channel feature extraction layers of the backbone and neck, reducing both computational and feature redundancy. Subsequently, a C2_FSAS module is designed to model frequency-domain relationships, enhancing long-range semantic dependency representation. An Efficient Channel Attention (ECA) mechanism is then applied to deep, high-level semantic features to enhance the weighting of salient feature channels. Finally, the traditional fixed interpolation upsample is replaced with a dynamic DySample upsample structure to recover fine edge features. Results indicate that Improved YOLOv11n-seg achieves segmentation accuracies of 97.0%, 98.1%, 99.2%, and 82.9% for P, R, mAP0.5, and mAP0.5:0.95, respectively. Compared with the original YOLOv11n-seg, the model achieves a 1.8% improvement in mAP0.5:0.95, reduces the number of parameters by 10.2%, and maintains a real-time inference speed of 34.8 FPS on the Jetson Xavier NX using TensorRT acceleration. Ablation studies validate the effectiveness of the four-module synergistic optimization, with C2_FSAS and DySample contributing most significantly to the mAP improvement. Moreover, the model exhibits enhanced edge segmentation and category discrimination capabilities. Overall, the results demonstrate that Improved YOLOv11n-seg achieves high accuracy and real-time performance, exhibits superior robustness compared to other models, and provides reliable technical support for intelligent impurity rate detection and edge deployment in mechanically harvested sugarcane.