An Image Classification Method Based on Multiple Visual Dictionaries

Abstract

In recent years, the bag of visual words (BoV) model is very popular in the field of computer vision. In particular, it has been widely used for image classification. The earliest method based on the BoV model, when constructing the visual dictionary, uses the clustering method to generate a single visual dictionary. This method does not consider the differences between categories of image dataset, leading to a poor performance of image classification. Later, an improved method is proposed that builds a visual dictionary for each category in the image dataset and then concatenates the visual dictionaries to produce the final visual dictionary. However, when there are many categories in the image dataset, using this method is time consuming. In addition, building a visual dictionary for each category leads to the visual words being too dense, losing the ability of distinguishing and expressing information and decreasing the performance of the image classification. In this paper, we proposed a new image classification method that is based on multiple visual dictionaries. This method addresses the above problems effectively and improves the performance of image classification significantly. To build a visual dictionary, the proposed image classification method first divides the image dataset containing many categories into N items randomly and uses the clustering method to cluster the local features of images of N items respectively. Next it generates a visual dictionary for each item in the image dataset. Finally it connects the visual dictionaries to produce the final visual dictionary of the entire image dataset. We conducted experiments on the caltech-101 and scene-67 image libraries and compared to the traditional methods. Our experimental results show that our proposed image classification method can match the best results published in the previous literatures in terms of classification accuracy rate.