Abstract: In most existing training data set for image annotation, keywords are usually associated with images instead of individual regions, so it is difficult to use supervised learning methods for image annotation. In this paper, based on quantum-behaved particle swarm optimization algorithm(QPSO) , a novel multi instance learning (MIL) algorithm was presented( QPSO-MIL)，we formulated image annotation as a supervised learning problem under Multiple-Instance Learning framework. This algorithm regards every image as a bag, and the feature vectors of the segmented regions in this image as instances. We defined a fitness vector for each particle based on the diversity density(Dl)) function. In the instance feature space we used QPSO to search the global maxima of DD function in each dimension simultaneously, and took the result as a concept point of the keyword,finally assigned corresponding key words to a test image according to the Bayesian maximum a posteriori probability criteria. Experimental results on ECCV 2002 data set indicated that the QPSO-MIL method is effective.

Key words: Multi-instance learning(MIL)，Image annotation,Quantum-behaved particle swarm optimization(QPSO)