基于树木整体图像和集成迁移学习的树种识别

为解决自然场景中拥有复杂背景的树木整体图像识别问题,提出了一种基于树木整体图像和集成迁移学习的树种识别方法。首先使用Alex Net、Vgg Net-16、Inception-V3及ResNet-50这4种在Image Net大规模数据集上预训练的模型对图像进行特征提取,然后迁移到目标树种数据集上,训练出4个不同的分类模型,最后通过相对多数投票法和加权平均法建立集成模型。构建了一个新的树种图像数据集——Trees Net,基于该数据集,设计了多类实验,并将该方法与传统的图像识别方法进行了分析比较。实验结果表明:该方法对复杂背景下树种图像识别准确率达到99. 15%,对于树木整体图像识别具有较好的效果。

Tree Species Recognition Based on Overall Tree Image and Ensemble of Transfer Learning

The automatic classification and recognition of tree image has important practical application value. Relevant research on traditional tree species recognition includes leaf recognition, flower recognition, bark texture recognition, and wood texture recognition. In order to solve the problem of recognizing the tree image with complex background in nature scenes, a tree species recognition method based on the overall tree image and ensemble of transfer learning was proposed. Four pre training models of AlexNet, VggNet-16, Inception-V3 and ResNet-50 were firstly used on ImageNet large scale datasets to extract features. They were then transferred to the target tree dataset to train four different classifiers. An ensemble model was finally established by the relative majority voting method and the weighted average method. A new tree image dataset called TreesNet was built and experiments were designed based on the dataset, including the comparative experiments of transfer learning and conventional methods.The experimental results showed that data augmentation can effectively solve the over fitting problem and the training model had better generalization ability and higher recognition rate. The image recognition accuracy of the tree species in the complex background with the method proposed reached 99.15%, which had a better effect on overall tree image recognition compared with the conventional classification methods of K nearest neighbor (KNN), support vector machine (SVM) and back propagation neural network (BP).