The Study Of Plant Leaf Image Recognition Based On Deep Learning Method

Plant species identification is very important for preserving biodiversity.Traditional plant recognition mainly relies on the subjective judgment of researchers,and this method is laborious.Therefore,this paper proposes two automatic plant leaf image recognition methods.I)Plant leaf image recognition method based on feature.Firstly,the plant leaf image is preprocessed,mainly including image graying,image binarization,median filtering and morphological filtering operations,and:finally the binary image and gray image for feature extraction are obtained.Then,the shape features are extracted on the binary image,and the texture features are extracted on the gray image.Since the extracted 19-dimensional blade features have different dimensions,the features are normalized,and then the principal component analysis method is adopted.The normalized features are dinension-reduced,and finally the reduced-dimensional features are input into the BP neural network classifier for training.Finally,the self-built dataset plant leaf images are used for classification and recognition,and the test set accuracy is 90.14%.2)Plant leaf image recognition method based on the transfer learning with convolutional neural network.First,each leaf image is randomly flipped horizontally and vertically to expand the dataset of leaf image.Then,the expanded dataset is divided into training set and test set at a ratio of 4:1.Then,AlexNet,VGG-16 and Inception-V3 models are pre-trained on the ImageNet large image dataset.Based on the theory of transfer learning,the learmed model parameters are transferred to the small sample plant leaf image dataset.On the basis of not changing other parameters,the number of neurons in the last full connected layer is replaced by the number of plant leaf species,so as to train the plant leaf recognition model.The methods are tested on the ICL dataset.The experimental results show that the accuracy of test sets obtained by AlexNet,VGG-16 and Inception-V3 pre-training lodel are 95.31%,93.86%and 95.40%,respectively.Finally,two methods of this paper are compared on the self-built plant leaf dataset.The experimental results show that the method based on transfer learning with convolutional neural network is more effective.