The Research On Bag-of-Words Based On Improved K-means And Hierarchical Cluster Algorithms

Bag of words(Bow) model was firstly applied to the documents classification. In recent years, the model has been widely used for its simpleness and efficiency with the further research of Bow. In addition, this model also plays an important role in image processing. Due to the importance of images in people’s daily life and work, how to find the expected information quickly and accurately from massive image databases is crucial. This paper has focused on image retrieval based on Bow model and it makes some improvement compared to the traditional Bow model.The main focus in this paper can be summarized three aspects as follows:1. This paper proposes a new clustering method combining K-means algorithm and hierarchical clustering algorithm. It firstly uses agglomerative hierarchical clustering algorithm to cluster the database and determine the cluster number and initial cluster centroids according to the effective evaluation criteria and maximum minimum distance algorithm. Then K-means algorithm is used for clustering again to obtain the final results. This paper chooses three databases from UCI real database for experiments to improve the effectiveness of the proposed method.2. This paper constructs a spatial visual bag of words model on the basis of Bow model. It divides the image into several blocks at different levels, so as to build the spatial visual bag of words model for later image processing. Then PCA algorithm and Bhattacharyya distance are used to evaluate the similarity between images and finish the image retrieval process. The experimental results demonstrate that the proposed method performs better and the results are closer to query intent of users.3. The relevance feedback technology is introduced to make the preliminary image retrieval results more satisfactory and obtain higher accuracy. According to the feedback marks given on image results by users, further study is done with relevance feedback technology based on feature weight. Then the feedback loop is made repeatedly until more satisfactory results are obtained. The experiments show the effectiveness of relevance feedback technology in image retrieval.