The Study Of Plant Leaf Identification Based On Deep Learning Method

Plant leaf recognition is important for new or scarce species identification and is important for improving the sustainable development of the pharmaceutical industry,balancing ecosystems and agricultural productivity.The effect of the traditional plant leaf recognition is depended on the “feature engineering”.It is necessary for human experts to participate in feature definition,feature extraction and feature processing.So that the problem of plant leaf identification is difficult in the natural environment.A method of automatic expression learning is proposed which extract features from 2-dimensional images,and used for classification and recognition of the plant leaves.We did the following:(1)Study the limitations of traditional feature engineering identification methods,Study the development of advanced learning and advantages.(2)Study the process of plant leaf identification based on characteristic engineering,including the extraction method of feature,the treatment and selection of feature,the design of classifier,and the problems of traditional characteristic engineering.(3)Study the process of plant leaf recognition based on deep learning,including image preprocessing with data augmentation in finite data set,key construction steps of feature extraction network,loss function selection and key construction steps of classification network.(4)Study the key issues of improving the recognition rate in the deep learning method.A deeper,wider,better,less parameters of the deep learning network construction methods.Faster training method and transfer learning usability and implementation method.In this paper,a method of plant leaf recognition is proposed,which establishes a deep convolution neural network model that can be used to extract features from 2-D images through automatic representation learning.Compared with the traditional feature engineering method,this method does not need human engineers to define?extract?process features.a method of plant leaf recognition is proposed which implements data augmentation to enhance robust of the model by horizontal mirror flip?lighting transformation?random cropping and de-averaging and normalization.Then,the preprocessed image is taken as the input of the 22-layer deep convolution neural network model.After a series of convolution and pooling operations,the robust feature map is formed.Finally,the feature map is scaled into one-dimensional eigenvector by global average pooling as the input of the full-connected layer,then the softmax loss function and the gradient descent algorithm are used to guide the learning and final classification.In this paper,"inception" structure is organized in the deep convolution neural network,which increases the depth and width,improves performance and decreases the parameters.In addition,in this paper,transfer learning based pre-training and fine-tuning are used,which can implements high performance classification and recognition model even if in the small data set with a short time.Experiments show that the plant leaf recognition method based on the deep learning has a good effect even for the natural environment plant leaves.This method achieves the 95.68% top1 recognition accuracy and 99.23% top5 recognition accuracy in 72 classes of CLEF2012.