LSA based multi-instance learning algorithm for image retrieval

Focusing on the problem of natural image retrieval, based on latent semantic analysis (LSA) and support vector machine (SVM), a novel multi-instance learning (MIL) algorithm is proposed, where a bag corresponds to an image and an instance corresponds to the low-level visual features of a segmented region. Firstly, in order to transform every bag into a single sample, a collection of “visual-word” is generated by k-means clustering method to construct a projection space, then a nonlinear mapping is defined using these “visual-word” to embed each bag as a point in the projection space, thereby obtaining every bag's projection feature. Secondly, the matrix consisted of all the projection features of training bags is regarded as a term-document matrix, and LSA method is used to obtain the latent semantic feature of each bag. As a result, the MIL problem is converted into a standard single instance learning (SIL) problem that can be solved directly by SVM method. Experimental results on the COREL data sets show that the proposed method, named LSASVM-MIL, is robust, and its performance is superior to other key existing MIL algorithms.