【摘要】 甘蔗茎节的快速准确识别对实现甘蔗田间作业和糖料产业精细化管理等任务具有重要意义，是田间产量测量的前提。为此，针对当前甘蔗人工测量成本高、人工参与量大和工作效率低等问题，设计了一种基于边缘计算的便携式甘蔗茎节智能识别系统。YOLOv5网络模型具有架构小、运行速度快等优势，通过对网络模型的进一步改进，使用K均值聚类方法调整锚框尺寸至最优，增强网络模型鲁棒性的同时也提高了茎节识别准确性；在卷积层后添加了CBAM注意力机制模块，使网络模型更加关注甘蔗茎节特征；引入VarifocalNet,使网络模型对于密集遮挡区域的甘蔗茎节有更好的识别效果。将改进后的算法移植入Jetson TX2嵌入式开发平台。试验结果表明：系统平均准确率(AP值)为89.89%,相比于之前的88.23%,提升了1.66%,单幅图像检测时间约为1.87s,视频检测速度可达15fps。系统可有效实现对甘蔗茎节的实时准确识别，为甘蔗田间产量测量提供了新的解决思路。更多还原

【Abstract】 Rapid and accurate identification of sugarcane characteristics is of great significance for sugarcane field yield measurement and fine management.Aiming at the problems of high production cost, large manual participation and low work efficiency in sugarcane field, a portable intelligent recognition system for sugarcane stem node based on edge computing was designed.Based on the advantages of small architecture and fast speed of YOLOv5 visual recognition algorithm, the improved algorithm is transplanted into the Jetson TX2 embedded platform.The multi-scale structure and K means clustering method are used to obtain the anchor frame size suitable for sugarcane features, which enhances the robustness of the network model and improves the accuracy of feature recognition. After the convolution layer, CBAM attention mechanism module is added, which includes channel attention module and spatial attention module. It can not only save parameters and computing power, but also ensure integration into the existing network architecture.The experimental results show that the average accuracy(AP value) of the system is increased by 1.66 %, the inference time of single image is about 1.87s, and the recognition rate of video is 15fps.The system can effectively realize the real-time and accurate recognition of sugarcane stem nodes, which provides a new solution for sugarcane field yield measurement.更多还原