Abstract: An image retrieval scheme based on multiscale analysis and SVM relevance feedback was proposed. First, a more accurate texture feature was extracted in Contourlet domain than Wavelet due to its multiresolution and directionality. One class and binary class SVM were combined to retrieve. The one class SVM can estimate the distribution of data in high dimensional space, and  exploit unlabeled data to get a primary similarity measure order. Then binary class SVM was used to get the labeled sample information through learning user's feedback, which finally improved the retrieval accuracy. The experimental results demonstrate the reasonability and effectiveness of the scheme. The most appropriate feedback image quantity and feedback times were proposed.

Key words: Contourlet, support vector machine, image retrieval