The Research And Implementation Of Image Classification Based On M-SVM And BOW

Along with the development of digital multimedia technique, a large number of images appeared and people’s demand for multimedia information rised quickly. Image classification technology has become a popular subject. This paper studied on the classification techniques of content-based images.The main work includes the followings:First, this article presented the background of image classification.In order to build the classification system, the paper discussed and analyzed the key technologies of image classification, including elaborating the principle and performance of common classification methods, image feature extraction and description and representation of local features. This paper also described naive Bayes classifier, adaboost classifier, K nearest neighbor and other commonly used classifiers, elaborated the principle of SVM classifier, compared the effect of different classifiers. This paper studied the description of the low-level image features and introduced SIFT operator, which has excellent descriptions and robustness, and, SURF operator, which has a faster calculation speed.Second, to overcome the semantic gap, the article proposed a new strategy to generate class-specific visual vocabularies by calculating the significance of the features, and designed a multiclass classifier based on the visual vocabularies generated before. The method generated the visual dictionaries by selecting and clustering the extracted SURF features using average significance. This method can take advantage of the image feature points those can easily lost in the traditional method, which is usually caused by the low quantity of the points. The method can also make the visual vocabularies smaller and increased the scalability of classification system.Third, this paper has built an image classification system to simulate the multiclass classifier based on class-specific visual dictionaries proposed, and achieved the result as expected.