Relevance Feedback Image Retrieval Algorithm Research Based On Support Vector Machine

Content-Based Image Retrieval(CBIR) approach offers effective solution to problems that the workload of manual annotation is large, great subjectivity is existed in annotation, few annotations are not enough to fully describe the complete information of image, annotations of the same image in different languages lead to different semantic understanding which are all brought from Text-Based Image Retrieval(TBIR) approach. By extracting visual features of images, choosing appropriate combination of features and query condition, the similar match of query image and retrieval images is achieved. Machine learning methods gradually reduce the difference between the semantic understanding of the image content from human people and computer, and improve the retrieval performance. By selecting image feature data as the input of support vector machine(SVM) model and setting the appropriate values of parameters in model, the classifier is built and used to classify the test samples and finally find the most similar image. In order to switch the center from computer to users, relevance feedback mechanism is proposed to mark retrieval results. The application of multiple feedback can make the retrieval results meet the purpose of users more precisely.Visual C++6.0 and MATLAB2013 a software are used as development tool. The color, texture and general features of image are extracted. Image block effectively adds the image space distribution information. Bag of words model can improve the classification accuracy. Different number of positive and negative image are used as the sample of feedback retrieval. SVM and extremely learning machine(ELM) are employed respectively to build classifier which are applied on image feature library. Simulations show that SVM has advantages on small-scale image feature library. For large-scale image feature library, SVM has some shortcomings and needs to be improved by combining with other algorithms.