Bag-of-Visual Words For Image Categorization

Content based image categorization, as a key technique of image retrieval, image recognition and image filtering, has become one of the most important research areas in the field of the pattern recognition. It aims at classifying images into different semantic categories. The bag of visual words model which has achieved a lot of success in image classification attracts more and more attention. However most existing approaches construct a visual vocabulary by simply clustering image regions represented with low-level visual features, where spatial context of image regions has not been well utilized. This thesis adopted and improved some methods which can take the spatial context into consideration. The main contributions of this thesis are as follows:Firstly, a new algorithm called evolving SOM-SD is proposed for the acceleration of the conventional SOM-SD for image classification. The most important advantage of the conventional SOM-SD is that it can deal with structural data and distinguish similar objects. However, it is not suitable for large database because of the extremely intensive computing task resulted from the consideration of the spatial information. We resolved the problem while keeping the capability of dealing with structural data by utilizing a hierarchical and evolving strategy. Experimental results demonstrated that our proposed method performs better than those without considering spatial context and can implement much faster than the conventional SOM-SD algorithm.Secondly, an algorithm called FCM-HS is proposed for the improvement of the FCM-S (FCM with spatial constraints). It's more robust to the noise compared with the k-means algorithm, and more efficient compared with the FCM-S algorithm. Experimental in the same environment showed a significant improvement in the robustness and efficiency in image classification.