Research On Image Classification Based On ELM And D-S Evidence Theory

With the development of the information age, image is widely used in the industry or on the Internet and daily life that a large number of images need to be processed every day, Whether the correct classification of each image is the basis of the application of the image in other areas.Image feature extraction is the basis of image classification, and a image classifier is designed according to the image feature vector. Now there are many studies on the image feature extraction and classification design and can achieve better results in practical applications. However, the current classification of images and design knowledge is not comprehensive enough, the image of some aspect of the expression should be improved, better image feature extraction methods have yet to be discovered. It is disadvantage to use in practical need to re-design the original classification model on use new features of the image can extract a deeper understanding of the novel features of the image design classifiers. To solve this problem, an image classification method is proposed from the perspective of classification model of scalability, the classifier does not need to be re-designed when a new feature vector is added and add the feature vector as evidence to classification models.This paper discusses an image classification design method, which is based on the combination of extreme learning machine and evidence theory:Firstly extract image texture features by contourlet transform and invariant moment with the method and gray level co-occurrence moments to, image color feature extraction based on the HSI color space model, a total of extracted the features of three groups respectively as the eigenvectors of the sample. Secondly the classifiers are designed based on the the three groups of characteristics, classification algorithm with ELM based on particle swarm optimization (PSO) limit, and puts forward a kind of method to obtain the basic probability assignment function that the output of the classification model as each kind of basic probability assignment function (also said to the extent of each type of support). Finally, the BPAF is fused with the theory of evidence to get the support degree of each kind of image. The method is modular and scalable, and can be used to obtain a set of basic probability assignment function by using the limited learning machine classifier when we get the new characteristic of the image. At last, Matlab is used to verify the validity of the image classification model.