Boosted Learning of Visual Word Weighting Factors for Bag-of-Features Based Medical Image Retrieval

In this paper, we investigate the bag-of-feature based medical image retrieval methods, which represent an image as a bag of local features, such as image patchs. To describe local feature bags effectively, most bag-of-feature methods learn a visual vocabulary containing a number of visual words via clustering. Different visual words have different importance in a medical image. The visual word weighting methods assign appropriate weights to the visual words to improve the performance of medical image retrieval. To improve bag-of-feature method, we propose an effective visual word weighting approach in this paper. We first analysis each visual word's discriminating power by modeling the sub distance function. This function compares only one single bin corresponding to the visual word. Then we treat each of them as a weak classifier for triplets and learn a strong classifier, using AdaBoost algorithm. The week classifiers' weights learned using boosting algorithm will be used as visual weighting factors. We carry our experiments on Image CLEF med 2007 and 2008 datasets. The vast experiments results show that our proposed methods have many advantages and work well for the bag-of-feature based medical image retrieval tasks.