基于卷积神经网络的小样本树皮图像识别方法

针对在树皮图像分类过程中图像训练数据数量少、识别准确率低的问题,提出一种基于卷积神经网络的小样本树皮图像识别方法。以5种常见树种的树皮图像作为研究对象,在基于卷积神经网络的Inception_v3模型基础上,对原始数据集进行数据增强的一系列操作,扩大数据集的数量;在此基础上,对所有数据集进行白化处理,以降低数据之间的冗余性,使得特征之间相关性较低;采用ReLU激励函数和Dropout方法,防止训练时引起的过拟合现象;同时,在模型的最后添加3层全连接层,增强模型的特征表达能力,采用softmax分类器。最终确定了一个10层CNN模型:5个卷积层、2个池化层、3个全连接层。结果表明,上述网络模型对数据集的识别准确率为94%,并且为验证本研究方法的可行性,分别在MNIST数据集、ImageNet数据集、CIFAR-10数据集进行测试,识别准确率分别为92%、90%、93%。因此,提出的方法在小样本的识别试验中具有较高的识别准确率和一定的可行性。

Small Sample Bark Image Recognition Method Based on Convolutional Neural Network

Aiming at solving the problems of low image training data and low recognition accuracy in bark image classification process,a small sample bark image recognition method based on convolutional neural network was proposed.The bark images of five common tree species were used as research objects.Based on the Inception_v3 model and convolutional neural network,a series of operations were performed on data enhancement of the original data set,to expand the number of data sets;on this basis,all data sets were whitened to reduce the redundancy between data,so that the correlation between features was decreased;ReLU excitation function and Dropout method were used to prevent over-fitting phenomenon during training.At the same time,three layers of fully connected layers were added at the end of the model to enhance the feature expression ability of the model,and the softmax classifier was used.A 10-layer CNN model was finalized:including 5 convolutional layers,2 pooled layers,and 3 fully connected layers.The experimental results showed that the above network model had a recognition accuracy of 94% for the data set.In order to verify the feasibility of the proposed method,the MNIST dataset,ImageNet dataset and CIFAR-10 dataset were tested,respectively.The recognition accuracy reached 92%,90%,and 93%,respectively.Therefore,the proposed method had higher recognition accuracy and certain feasibility in the identification test of small samples.