高级检索+

基于深度学习的树种识别系统设计与试验

Design and Experiment of Tree Species Recognition System Based on Deep Learning

  • 摘要: 为解决目前树种识别任务中存在的准确率低、工作强度大、成本高和效率低的问题,基于深度学习方法,利用机器视觉软件HALCON,结合Arduino程序开发技术,设计并开发林木树种识别系统。通过搭建林木树种识别系统主体硬件设备,自动采集6种林木的3 000张图像,并按7∶1.5∶1.5的比例随机划分为训练集、验证集和测试集;基于HALCON目标检测框架,比较分析3种SqueezeNet、Inception-V3和ResNet-50目标检测特征提取网络的训练结果和评估结果,选择性能表现最好的SqueezeNet网络模型,试验分析不同环境和生长状况的树种对其影响,构建并设计林木树种识别系统。系统识别试验结果表明,模型的平均精度达0.735,在测试集上识别准确率和召回率均高于99.5%;可在不同光照条件下准确识别不同直立程度、胸径大小的林木,具有较强的泛化性能;田间试验随机测试结果表明,网络模型准确率均高于93%,同时系统具有良好的稳定性和实用性,可满足作业实际工况的要求。该研究为林木资源智能化管理提供技术支撑。

     

    Abstract: To solve the low accuracy, high work intensity, high cost, and low efficiency problems in the tree species recognition, a tree species recognition system was designed and developed based on the deep learning algorism with using the machine vision software HALCON and Arduino program development technology. By building main hardware equipment of the tree species recognition system, 3 000 images of six species trees were automatically collected. All the images were randomly divided into the training set, verification set, and test set as the ratio of 7∶1.5∶1.5. Based on the HALCON target detection framework, this research compared and analyzed the training results and evaluation results of three target detection feature extraction networks: SqueezeNet, Inception-v3, and ResNet-50, respectively. The SqueezeNet network model was chosen due to its good performance to experiment the impact of tree species in different environments and growth conditions, and then a tree species recognition system was designed and constructed. The recognition experimental results showed that the average accuracy of the model was 0.735, and the recognition accuracy and recall rate on the test set was higher than 99.5%. It could accurately identify different species of trees with different upright degrees and DBH under different light conditions and the system had strong generalization performance. The on-site random test results showed that the accuracy of the network model was higher than 93%, and the system had good system stability and practicability, which can meet the requirements of actual working conditions. This research provides technical support for the intelligent management of forest resources.

     

/

返回文章
返回