The content of impurities in mechanically harvested sugarcane is a critical factor for evaluating harvest quality and determining market price. To enable intelligent detection of impurities in mechanically harvested sugarcane, this study proposes an impurity detection method based on an improved YOLOv11n-seg model. The method integrates four enhancement modules into the original YOLOv11n-seg architecture. Firstly, a lightweight C2_Ghost module is introduced into the high-channel feature extraction stages of both the backbone and neck, thereby reducing computational complexity and feature redundancy. Subsequently, a C2_FSAS module is designed to perform frequency-domain relationship modelling, enhancing long-range semantic dependency representation. An Efficient Channel Attention (ECA) mechanism is then applied to deep high-level semantic features to adaptively reweight salient feature channels. Finally, the traditional fixed interpolation-based upsampling operation is replaced with a dynamic DySample upsampling strategy to recover fine-grained edge features. Experimental results indicate that Improved YOLOv11n-seg achieves segmentation performance of 97.0%, 98.1%, 99.2%, and 82.9% in terms of P, R, mAP _0.5 , and mAP _0.5:0.95 , respectively. Compared with the original YOLOv11n-seg, the proposed model achieves a 1.8% improvement in mAP0.5:0.95, a 10.2% reduction in parameter count, and maintains a real-time inference speed of 34.8 FPS on the Jetson Xavier NX under TensorRT acceleration. Ablation studies validate the effectiveness of the four-module synergistic design, with C2_FSAS and DySample contributing most significantly to the improvement in mAP. Moreover, the model exhibits enhanced edge delineation accuracy and inter-class discrimination capability. In summary, the Improved YOLOv11n-seg achieves a favourable balance between segmentation accuracy and real-time performance, enabling precise segmentation of sugarcane segments and diverse impurity types. The proposed method provides reliable technical support for intelligent impurity rate detection in mechanically harvested sugarcane and practical deployment on edge computing platforms.

China is the world’s third-largest sugarcane producer. When mechanically harvested sugarcane enters the sugar mill, impurity rate detection is required. However, due to the piling up of sugarcane, significant errors may occur in the detection results. Therefore, this research addresses the issue of low accuracy in machine vision detection due to the dense stacking of sugarcane. An innovative graded device was developed, featuring a three-stage progressive geometric constraint system with roller-belt gaps of 100 mm, 45 mm, and 33 mm, alongside differential traction with speed ratios of 3:1, 4:1, and 5:1. Utilizing the normal distribution characteristic for the diameter of 500 sugarcane stalks, the gap parameters were refined through a dynamic stiffness model. Through power validation and multi-factor orthogonal experiments, the study uncovered the interactive influences of sugarcane weight, primary conveyor belt speed, and speed ratio on the single-layer rate and area ratio. Our findings indicate that sugarcane weight is the primary determinant of the material’s single-layer rate, while the speed ratio is crucial for managing sugarcane distribution density, more so than the primary conveyor belt speed. Notably, increasing the speed ratio from 3:1 to 5:1 results in a decrease in area ratio from 26.8% to 22.0%. After utilizing the graded differential device, the average accuracy of machine vision detection achieved 94.90%, with only two misidentifications on average. In comparison to the control group, detection accuracy improved by 26.93%, misidentifications dropped by about 81.80%, and detection speed was recorded at 55.5 ms. These outcomes confirm that the device not only enhances detection accuracy but also significantly lowers the misidentification rate, thereby creating a stable, clear, and efficient detection environment.