A Scene Recognition Algorithm Based On Multi-Instance Learning

In Bag of Words image presentation model, visual words are generated by unsupervised clustering, which leaves out the spatial relations between words and results in such shorting comings as limited semantic description and weak discrimination. To solve this problem, we propose to substitute visual words by visual phrases in this article. Visual phrases built according to spatial relations between words are semantic distrainable, and they can improve the accuracy of Bag of Words model. Considering the traditional classification method based on Bag of Words model is vulnerable to the background, block and scalar variance of an image, we propose in this article a multiple visual words learning method for image classification, which is based on the concept of visual phrases combined with Multiple Instance Learning. The final classification model is able to show the spatial features of image classes. Experiments performed on standard image testing sets, Caltech 101 and Scene 15, show the satisfying performance of this algorithm.