Plant Leaf Hybrid Classification Algorithm Based On Convolutional Neural Network

There are many different kinds of plants on earth,andto recognition and classification of plant speciesis veryimportant for plant protection,improved variety identification and pest protection.Leafwhich is an important organ of plant,has the advantages of long retention time,convenient collectionand the 2D structure for image storage.So leaf has become an important basis for classification and identification of plant species based on computer digital image processing.The traditional classification of plant leaf generally needs to manually extract the feature of the leaf,excellent design feature extractor and repeatedly adjust the parameters to achieve a satisfactory accuracy.However,the performance of these methods often depends on the rationality of artificial feature selection,and this artificial feature has a great deal of blindness,and does not have versatility and mobility.In order to solve the above problems,thethesisselects a hybrid CNN-SVMalgorithm.First,theresearch choose Convolutional Neural Network(CNN),it can be used in two dimensional images directly,automatically extracted higher-level features of leaf.Then,using Support Vector Machine(SVM)learns these higher-level features,classifythese plant species.Then,thethesis selects the Flavia,ICL and Leaf100 three public leaf data set as the research object,using hybrid CNN-SVM method on the three leaf data sets to verify its effectiveness.The main research contents and results are shown as follows:1)A pre-processing of minimum enclosing circle has been usedto ensure the centralized location of the leaf and normalization size of leaf image,this pre-processing method provide basic data for further Convolutional Neural Network(CNN).Thethesisuses many pre-processing methods such as graying,binarizationand contour extraction of the largest closed area,minimum enclosing circle and bilinear interpolationtoremove the noise and interference elements in the leaf images and ensure thecentralization of the leaflocation and the normalization of leaf imagesize(prevent the shape of leaf in the image which is inconsistent width and height changes significantly during the image scaling transformation),which can provide basic data for subsequent further Convolutional Neural Network(CNN)to extract higher-level features of leaf.2)Transfer learning with Convolutional Neural Network(CNN)has been designed to extract higher-level features of leaf,next the output of level having most optimal features has been selected for the input of further classifier.Compared the methods between an entire CNN and the transfer learning with CNN,the training results showed that transfer learning with Convolutional Neural Network(CNN)can accelerate the convergence speed and improve the recognition rate.Then,take advantage ofextracting higher-level features of leaf in Convolutional Neural Network(CNN),and select the Leaf100 leaf data set as the research object,visualizingand training the output results of each layer features,concluding the level having most optimal features.3)Plant leaf hybrid classification algorithm based onConvolutional Neural Network(CNN)has been used,compared the results of experiments,andit proves that the hybrid CNN-SVM is effective onplant leaf classification.Mixing the Convolutional Neural Network(CNN)withSupport Vector Machine,k-Nearest Neighborand Random Forestseparately,the results shows that the performance of hybrid CNN-SVM is best.Then,selecting hybrid CNN-SVM to make more detailed and comparative experiments,which demonstrate hybrid CNN-SVM haves effectiveness and generality in three leaf data sets,specially,recognition rate of leaf reaches 99% whentraining data reaches a certain size on Flavia leaf data set.At the same time,it demonstratesthe shape of leaf plays a very important role in leaf classification.