基于GAN-DCNN的树叶识别  

Leaf Identification Based on GAN-DCNN

作　　者：徐竞怡 张志 闫飞[1] 张雯悦 Xu Jingyi;Zhang Zhi;Yan Fei;Zhang Wenyue(Beijing Key Laboratory of Precision Forestry,Beijing Forestry University,Beijing 100083;State Key Laboratory of Resources and Environmental Information System Institute of Geographic Sciences and Natural Resources Research Chinese Academy of Sciences,Beijing 100101;University of Chinese Academy of Sciences,Beijing 101408;SenseTime Research,Beijing 100039)

机构地区：[1]北京林业大学精准林业北京市重点实验室,北京100083 [2]中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室,北京100101 [3]中国科学院大学,北京101408 [4]商汤科技研究院,北京100039

出　　处：《林业科学》2024年第4期40-51,共12页Scientia Silvae Sinicae

基　　金：北京林业大学中央高校基本科研业务费专项资金项目(2017PT07);西藏自治区科技厅中央引导地方项目(XZ202301YD0043C);北京林业大学2021年大学生创新训练项目(X202110022004)。

摘　　要：【目的】利用深度学习进行树叶识别时需要大量训练样本,当样本量不足、图像风格单一会导致识别准确率不稳定。研究利用少量的样本进行树叶图像增殖和风格转换,可极大减轻数据采集的负担,为提升林业调查信息化、智能化提供有效的技术手段和理论支撑。【方法】采集6种树种的树叶图像建立数据集,引入light-weight GAN对图像进行增殖和风格转换,扩充人工拍摄的树叶数据集,通过在该数据集与原数据集上分别应用AlexNet、GoogLeNet、ResNet34和ShuffleNetV2四种深度卷积神经网络进行训练,分析生成对抗网络的图像增殖技术在树叶识别中的作用。综合模型准确率和训练时间等性能指标选择最优模型,同时对模型的学习率进行调整。使用测试样本对参数优化后的模型进行验证,分析该方法在实践中的可行性和意义。【结果】基于生成对抗网络生成的样本具有高清晰度,高保真性,能够有效地辅助神经网络模型的训练工作,同时也丰富了样本类别,使之获得包含更多不同季节、形状、健康状况的树叶图像。与原始数据集相比,AlexNet、GoogLeNet、ResNet34和ShuffleNetV2四种网络在新数据集的训练上均表现出训练误差更小、验证精度更高的特点,其中学习率为0.01的ShuffleNetV2模型对该数据集的训练效果最好,训练时最高验证精度为99.7%。使用未参与训练的测试样本对该模型进行验证,模型对各树叶的识别效果较好,模型的总体识别准确率高达99.8%。与未使用GAN技术的普通深度卷积神经网络相比,本文提出的模型对树叶识别准确率明显提升。【结论】生成对抗网络可以有效地扩充图像数量,对图像进行风格转换,与深度卷积神经网络相结合,可以显著提高树叶识别准确率,适合应用于林业树叶识别领域。【Objective】A large number of training samples are needed for leaf identification based on deep learning.Insufficient sample size and single image style can affect the identification accuracy.However,studying the use of a small number of samples for leaf image reproduction and style transformation can greatly reduce the burden of data collection,which would provide effective technical means and theoretical support for improving forestry survey informationize and intelligence.【Method】The leaf images of 6 tree species were collected to establish a dataset,and the light-weight generating adversarial networks(GAN)was introduced to propagate images and style translation,and expand the manually shot leaf dataset.Four deep convolutional neural networks(DCNN),AlexNet,GoogLeNet,ResNet34 and ShuffleNetV2,were applied to train the dataset and the original dataset,respectively,by which the role of image augmentation techniques of GAN in leaf recognition was analyzed.The optimal model was selected based on performance indicators such as model accuracy and training time,and the learning rate was adjusted.Finally,the test samples were used to verify the optimized model,and the feasibility and significance of the method in practice were analyzed.【Result】The samples with high definition and high fidelity were generated based on generative adversarial networks,which was able to effectively enrich the sample category,and obtain leaf images with different shapes,and health conditions at different seasons.Compared with the original dataset,AlexNet,GoogLeNet,ResNet34 and ShuffleNetV2 all showed smaller training errors and higher validating accuracy on new dataset during the model training.Among them,the ShuffleNetV2 model with a learning rate of 0.01 had the best training effect on this dataset,whose highest validating accuracy was 99.7%.The model was verified by using the test samples and a good recognition performance on each leaf was achieved,and the overall recognition accuracy of the model was up to 99.8%.Compared with

关 键 词：树叶识别 生成对抗网络 深度卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]