基于YOLO v8-SD的预切式甘蔗种质量检测方法研究

Quality Inspection Method of Pre-cut Sugarcane Seed Based on YOLO v8-SD

摘要:

甘蔗种质量检测是甘蔗预切备种过程中的关键环节。针对实际备种场景中环境复杂、边缘设备计算资源受限的问题，提出一种基于改进YOLO v8n的轻量化甘蔗种质量检测方法。通过引入混合局部通道注意力（Mixed local channel attention，MLCA）和部分卷积(PConv)改进网络结构，有效减少模型的冗余计算并增强对关键特征的关注能力；在颈部网络使用DySample动态上采样，以更准确地获取特征点的边界和细节信息；此外，采用重参数化的共享卷积结构对检测头进行改进，进一步减少模型的复杂度；最终形成了甘蔗种质量检测模型YOLO v8-SD，将其结合单目相机测距算法，实现蔗种切口与相邻蔗节的距离测算。试验表明，YOLO v8-SD模型的mAP50达98.3%，帧检测速率为142.9f/s，模型内存占用量为3.45MB，相较原始模型参数量和浮点运算量分别减少47.8%和33.3%，模型内存占用量减小41.9%，帧率提升7.8f/s。此外，使用本文方法的切口距离测算平均相对误差小于6.1%。将改进后的模型部署在NVIDIA Jetson Orin NX开发板上配合甘蔗种筛选系统进行样机试验，试验结果表明，对于不同品种甘蔗种质量检测的平均准确率为97.33%，满足实际应用需求。

Abstract:

Quality detection of sugarcane seed segments is a key step in the pre-cutting preparation process. To address challenges such as complex environments and limited computational resources on edge devices in practical seed preparation scenarios, a lightweight sugarcane seed quality detection method was proposed based on an improved YOLO v8n model. By introducing mixed local channel attention (MLCA) and partial convolution (PConv) to the network structure, redundant computations were effectively reduced while enhancing the model’s focus on critical features. A dynamic upsampling module (DySample) was adopted in the neck network to more accurately capture the boundaries and details of feature points. Additionally, the detection head was improved by using a re-parameterized shared convolution structure, further reducing model complexity. The final model, named YOLO v8-SD, was integrated with a monocular camera distance estimation algorithm to calculate the distance between the cut surface and adjacent sugarcane nodes. Experimental results showed that YOLO v8-SD achieved an mAP50 of 98.3%, a frame detection speed of 142.9f/s, and a model size of only 3.45MB. Compared with the original model, the parameter count and FLOPs were reduced by 47.8% and 33.3%, respectively, the model size was reduced by 41.9%, and the frame rate was increased by 7.8f/s. Furthermore, the proposed method achieved an average relative error of less than 6.1% in cut-surface distance estimation. The improved model was deployed on an NVIDIA Jetson Orin NX development board and tested in a prototype sugarcane seed screening system. The results demonstrated an average detection accuracy of 97.33% across different sugarcane varieties, meeting the requirements of practical applications.