Classification Of Natural Images Based On CNN-SVM And Algorithm Optimization

With the development of society,especially the popularity of smart phones,people get images more and more convenient and fast.Therefore,more and more image data are stored on the network,which not only brings great difficulties to the storage of image data,but also enriches the information stored in image data.It is becoming more and more important to obtain the information in image.It is impossible to complete timely archiving,organization and management of massive image data only by human resources.How to use so many image information and locate the interested material is a great challenge to the classification technology of image information.Image classification is a basic work.It is widely used not only in image classification management and information extraction,but also in target recognition,face recognition and image retrieval.It has important value and significance in other research fields.With the continuous development of information technology,people will use computer information technology in a wider range of fields.With the idea of"Precision Forestry",researchers began to pursue the so-called ecological management informationization,and image classification technology gradually began to be used in natural images.Images taken in natural scenes contain abundant information and are disturbed by environment,light,weather and noise.Therefore,how to classify the objects we are interested in from the images taken in natural scenes has become a complex and exploratory work.Support-Vector-Machine(SVM)was officially published in 1995.Because of its excellent performance in text categorization tasks,it soon became the mainstream technology of machine learning,and directly set off the climax of statistical learning around 2000.This technology has always been a research hotspot in the international machine learning field.Many scholars have introduced it into image classification and achieved good results.At present,many researchers have made a variety of improvements to SVM so that it can be applied to different scenarios.This thesis is based on the improved CNN-SVM which combines the popular Convolutional Neural Network with SVM.This thesis starts from the research significance of image classification.Firstly,the research status of image classification is introduced,including image classification technology,image feature extraction,feature coding technology,machine learning algorithm used in image classification.The ideas and principles of these technologies are introduced.Then it introduces some theories involved in the experiment,such as the principle of SVM classification,the steps followed in image classification,how to extract image features,direction gradient histogram and gray level co-occurrence matrix,etc.On the basis of related theories and algorithms,focusing on natural images,this thesis analyses the problems faced by natural image classification under complex background,and proposes a CNN-SVM-based natural image classification algorithm.Classifier model integrates traditional CNN network with SVM classifier reasonably,uses hidden layer of CNN network to extract features automatically,and uses SVM classifier with better classification effect to classify.At the same time,considering that some strong features may be omitted from CNN feature extraction,this thesis proposes to extract strong features manually and input them into SVM classifier after fusion with CNN features.This experiment takes typical representative tree images of natural images as experimental materials,and will experiment on the classification of tree images in MATLAB.Before using CNrN-SVM to classify natural images,a large number of representative data are extracted from the data to be processed as training samples,and then two image features,directional gradient histogram and gray level co-occurrence matrix,are selected to analyze the image to be tested.480 images of four kinds of natural trees are selected from the online search engine,and they are composed into image database,which includes 120 pictures of maidenhair tree,maple,pine,willow respectively.40 representative pieces of each type were selected as training data,and the remaining 80 pieces were used as test data.Finally,through the analysis of the first experimental results,we find out some reasons for the unsatisfactory classification results,and propose an improved CNN-SVM classifier.Then we do an experiment and compare the experimental results with the previous ones,and draw the final conclusion.Finally,the experimental results are evaluated and prospected.