A novel multi-instance learning algorithm with application to image classification

Image classification is an important research topic due to its potential impact on both image processing and understanding. However, due to the inherent ambiguity of image-keyword mapping, this task becomes a challenge. From the perspective of machine learning, image classification task fits the multi-instance learning (MIL) framework very well owing to the fact that a specific keyword is often relevant to an object in an image rather than the entire image. In this paper, we propose a novel MIL algorithm to address image classification task. First, a new instance prototype extraction method is proposed to construct projection space for each keyword. Then, each training sample is mapped to this potential projection space as a point, which converts the MIL problem into standard supervised learning problem. Finally, an SVM is trained for each keyword. The experimental results on a benchmark data set Corel5k demonstrate that the new instance prototype extraction method can result in more reliable instance prototypes and faster running time, and the proposed MIL approach outperforms some state-of-the-art MIL algorithms.