Bag-of-Words For Image Categorization

With the development of computer and multimedia technology, digital image is widely used in various fields in recent years. It replaces traditional text-based recording mode because it’s easy to understand. We need to classify images to get useful information quickly based on its context when facing such a large number of images. The traditional artificial tagging classification is very inefficient. So it is replaced slowly by computer classification system.BOW (Bag of Words) model was used for text analysis before and transfer to the image now. BOW model replace image by image feature combinations to classify images. This algorithm ignores some local features and space contact, so it reduces the classification results. We add improved spatial pyramid matching to the component-based bag of words model, and this method can extract local f eatures with stronger discriminations than the visual vocabulary collections of the bag of words model. The algorithm improved are listed in two aspects as folio(1)The first improvement is spatial pyramid matching to image’s words, and count words frequency histogram. Spatial pyramid matching method use multi-level and multi-granularity grids to process words. Bag of words have some image space information and local detail features that its discrimination is better after processing.(2) Image’s feature is handled second time by Histogram intersection after firstly processing by spatial pyramid matching. Through histogram intersection method is very simple, but it very useful. After this process, the same class image is more similar and the different class images are away from each others.Each point is assigned to the centroid which has minimum weighted distance while clustering features. This algorithm enhances the clustering quality by reducing the probability of misclassifying the points of big sparse clusters to its neighboring compact clusters.We did image classification experiments with component-based bag of words model and improved spatial pyramid matching on CaltechlOl image database. Experimental results demonstrate that this approach effectively overcomes the shortcomings of traditional component-based bag of words model. It has a better image classification performance. Component-based bag of words model and improved spatial pyramid matching method enhances classification accuracy about 7%. This method increase the lower classification accuracy to enhance the whole system’s classification accuracy.