Interactive metric learning system for similar image search using Linear Discriminant Analysis

Similar image search algorithm is important technique for efficiently search the images of our interests in the large pool of image data. In this paper, we propose an interactive metric learning search system using Linear Discriminant Analysis (LDA). We rank the search result based on the Mahalanobis distance calculated from SIFT feature vectors and use a LDA which enables users to update a search result until the returned results are relevant or satisfied by users. Moreover, we also examine the efficient learning rate and dimension size of feature vector for this model to enhance the search results. The efficiency of our model is confirmed through the intensive experiment using dataset from Caltech-256 which shows that the third updated result's accuracy can be increased by more than 57% from the default result's accuracy using our method.