For the problem of object-based image retrieval, a novel semi-supervised multi-instance learning (MIL) algorithm
is presented. This algorithm regards the whole image as a bag, and low-level visual feature of the segmented regions as
instances. Then according to the principle of maximum “local density”, a collect of “visual word”is generated to construct
a projection space. A nonlinear function is defined by using these “visual word” to transform each bag into a point in
the projection space, and extracting all bags’ projection features. Rough set (RS) method is used to reduce the redundant
information in the projection feature, then semi-supervised transductive support vector machine (TSVM) algorithm is applied
to train classifiers. Experimental results on the SIVAL image set show that this algorithm is feasible, and the performance is
superior to other MIL algorithms.