Research On Image Classification Based On Weighted Multi-features Fusion And SVM

Content-based image classification mainly depends on the extraction of low-level feature and the structure of multiple classifiers. SVM (Support Vector Machine, SVM) is known as a high-efficiency classifier with good ability of supervised learning and high generalization.Researches on image classification based on SVM become the theoretical foundation of image classification for many scholars.In this paper we mainly study the structure of multiple classifiers and Multi-features fusion to train the SVM classifier for a higher accuracy.In this paper we respectively make a simple introduction to the extraction algorithm of image content features and then choose better robustness extraction algorithm.To extract color, this paper choose extraction algorithm based on RGB color statistics, respectively extract the color component histogram of R, G, B.To extract the texture, this paper choose Gray-level Co-occurrence Matrix to indicate texture feature.Based on shape feature extraction and selection for Canny edge detection operator to extract the edge contour, combining with the method of invariant moments to describe the shape feature.This paper studies the structure method of multiple classifiers and puts forward the improved binary tree SVM method of multiple classifiers with making improvement on the basis of the binary tree SVM classifier.By calculating the Euclidean distance between image feature vector we can measure the similarity of two classes of image and then choose sorting order according to similarity.This paper chooses low similarity as priority order for classification to avoid the binary tree SVM classification disconnect two classes only to the negative side because of the less number of positive side.In this paper, we compare the original binary tree SVM classifier and improved classifier by experiment and verify the validity of the improved multiple classifier.This paper presents a weighted fusion method based on research on the characteristics of the weighted fusion method to fuse the feature vector,he proposed fusion method is based on class is the number of correct classification,which makes the fusion features as the input vector of SVM and perform an experiment based on the improved multiple classifiers as follows:for each category in data set, compared the higher accuracy based on multiple features fusion with the higher accuracy based on single feature to prove that the method based on feature fusion has a better classification result.In this paper we compare the various classes based on different methods to classify for the optimal results, feature fusion method of one kind is based on the numbers of correct classification when based on single feature.Because of various sorts, the class number is less than the negative number, so here we conbine classification accuracy and the numbers of correct classification as our contrast benchmark.By comparing the experimental results the feature fusion method is proved effective.