Abstract: In order to solve the problem that the speed of sugarcane bud recognition does not match in the directional mechanized planting of sugarcane seeds, this paper proposes a rapid recognition method of sugarcane buds based on the improved YOLOv4-Tiny. The backbone network is integrated with the SE-Resnet module to realize the attention mechanism to enhance the characteristics of sugarcane buds. The neck network structure adds an additional layer of prediction scale and performs Kmeans reclustering of the anchor box, and uses the detailed information of the shallow network to improve the performance. The detection ability of the model to small target sugarcane buds, the design of humancomputer interaction interface displays the sugarcane bud identification and positioning information in real time. In this study, the improved YOLOv4-Tiny algorithm and NCS2 accelerated reasoning will be deployed in the Raspberry Pi 4B device for testing. The test results show that the recognition accuracy of sugarcane buds is 95.87%, and the average precision and mean mAP is 92.46%. Based on the Raspberry Pi, the detection speed is 0.61 s, and the model size is only 23.2 MB. It realizes accurate and fast identification of sugarcane buds deployed in embedded devices, solves the problem that the slow recognition speed of sugarcane seeds restricts the speed of mechanized sowing of sugarcane seeds, and provides a solution for mechanized directional planting of sugarcane seeds.

Key words: sugarcane sprout identification, YOLOv4-Tiny, Raspberry Pi 4B, SE-Resnet, Kmeans reclustering

摘要： 为解决蔗芽识别在蔗种定向机械化种植速度不匹配问题，提出基于改进YOLOv4-Tiny的蔗芽快速识别方法。将主干网络融合SE-Resnet模块实现注意力机制以增强蔗芽的特征；颈部网络结构多增加一层预测尺度并进行锚框的Kmeans重聚类，利用浅层网络的细节信息来提高模型对小目标蔗芽的检测能力；设计人机交互界面实时显示蔗芽识别定位信息。该研究将改进YOLOv4-Tiny算法和NCS2加速推理部署在树莓派4B设备中测试，试验结果表明：识别蔗芽精度达到95.87%，平均精度均值mAP为92.46%，基于树莓派检测速度为0.61s，模型大小仅为23.2MB。实现部署在嵌入式设备中准确快速识别蔗芽，解决蔗种蔗芽识别速度慢制约蔗种机械化播种速度问题，为蔗种机械化定向种植提供解决方案。

关键词: 蔗芽识别, YOLOv4-Tiny, 树莓派4B, SE-Resnet, Kmeans重聚类