改进YOLOv3网络提高甘蔗茎节实时动态识别效率

为推广甘蔗预切种良种、良法种植技术,结合甘蔗预切种智能横向切种机的开发,实现甘蔗切种装置对蔗种特征的连续、动态智能识别。该文通过甘蔗切种机黑箱部分内置的摄像机连续、动态采集整根甘蔗表面数据,采用改进的YOLOv3网络,建立智能识别卷积神经网络模型,通过拍摄装置内部的摄像头对输入识别系统的整根甘蔗的茎节图像特征进行实时定位与识别,并比对识别信息,及时更新茎节数据,识别、标记出茎节位置,再经过数据处理得到实时的茎节信息,输送到多刀数控切割台进行实时切割。经过训练及试验测试,结果表明:经过训练及试验测试,模型对茎节的识别的准确率为96.89%,召回率为90.64%,识别平均精度为90.38%,平均识别时间为28.7 ms,与原始网络相比平均精确度提升2.26个百分点,准确率降低0.61个百分点,召回率提高2.33个百分点,识别时间缩短22.8 ms,实现了甘蔗蔗种的连续、实时动态识别,为甘蔗预切种智能横向切种机的开发提供数据基础。

Increasing the real-time dynamic identification rate of sugarcane nodes by improved YOLOv3 network

To popularize the technology of sugarcane pre-cutting seed, and cultivation with good method, combine with the development of intelligent transverse sugarcane pre-cutting seed-cutting machine, and realize the continuous and dynamic intelligent recognition of sugarcane seed characteristics by sugarcane seed-cutting device, in this study, an intelligent recognition convolution neural network model based on improved YOLOv3 network was established by continuously and dynamically collecting the surface data of the whole sugarcane through the camera built in the black box of the sugarcane cutting machine. The real-time location and recognition of the image features of the whole sugarcane cane nodes in the input recognition system was carried out by the camera inside the system. Compared with the recognition information, the improved network timely updated the sugarcane nodes data, identified and marked the position of the sugarcane nodes, and then got real-time sugarcane nodes information through data processing, which was transmitted to the multi-tool cutting table for real-time cutting. In this paper, based on the improved YOLOv3 network, a sugarcane nodes recognition system was established. The image acquisition was carried out by the camera in the sugarcane cutting system. The video data of sugarcane was collected before the training of this network, and then the image data was processed to establish training set, validation set and test set. The training data sets of different improved models were tested and the best model was selected as the model in this paper. Through the training and test, the measured results showed that the recognition accuracy of the model for sugarcane nodes was 96.89%, the recall rate was 90.64%, average recognition accuracy AP was 90.38%, and the average recognition time of pictures was 28.7 ms. Compared with the original network, the AP was improved by 2.26 percentage point, the accuracy was decreased by 0.61 percentage point, and the recognition time was shortened by 22.8 ms. At present, the recognition of sugarcane nodes still remained in single or basic image processing and recognition, and there was still a lack of fast processing methods for the whole sugarcane image. In this study, we proposed to use the improved YOLOv3 network to recognize and locate sugarcane features, and to establish the recognition model of sugarcane nodes through network training. On the basis of the accuracy and speed of the original network identification and location, the speed of identification, detection and the recognition rate were further improved. The whole sugarcane can be identified and processed quickly in real time, which can meet the needs of various sugarcane seed cutting. Combining with the other parts of the intelligent sugarcane cutting machine system designed by our research group, the whole cutting process can be mechanized and intellectualized, which can greatly improve the quality of sugarcane cutting, reduce the labor intensity and time, and greatly improve the production efficiency. It provides a research basis for the industrialized production of sugarcane pre-cutting and realizes the sugarcane production. Continuous and real-time dynamic identification of sugarcane seeds lays the application foundation for the development of intelligent transverse cutting machine for sugarcane pre-cutting.