Study On The Application Of The Improved K-means Clustering Algorithm In Image Retrieval

Traditional content-based image retrieval uses the sequential search, however, for massive, high-dimensional image data, this retrieval method has clearly inadequate in efficiency. It becomes more and more important to do necessary pre-processing and establish index for the image database. So we introduce efficient k-means clustering algorithm and the cohesion Hierarchical Clustering into the pre-processing of the Image database in order to establish the hierarchical index structure, which could achieve the purpose of fast retrieval. In the pre-processing, we improve the k-means clustering algorithm to determine the k values and the initial cluster centers, and give an adaptive clustering algorithm, to ensure a large degree of similarity within the same class, but little similarity between different classes. Our main work is as follows:1. Propose a weighted histogram of color image color feature representation based on the Delaunay triangulation. The feature preserves the advantages of the traditional color histogram feature that is simple and effective and not sensitive to changes in right perspective, in addition, it composite the color space distributed information into the representation of histogram, which could improve the right search rate; In selecting the color model, according to HSV Color model features, it can be divided into black, white and color three areas, further improving the accuracy of the color model.2. Propose an adaptive k-means clustering algorithm. The clustering algorithm uses norm of the Davies-Bouldin Index to determine the best number of clusters for improving the maximum minimum distance method to select the initial cluster centers, effectively solve the problems that k value is difficult to determine, and improper choice of initial centers cause instability clustering results.3. Establish the hierarchical index structure based on the adaptive k-means clustering and hierarchical clustering. Use the clustering algorithm to cluster the Image features data, and according to the visual characteristics similarity of the characteristics library to organize the hierarchical indexing structure. By the index structure, it reduces the number of image feature database access and access data, to achieve fast image retrieval.4. Based on the above theory, design and implement a content-based image retrieval experiment platform. Through the experimental data, it proves that our method that based on hierarchical index structure in content-based image retrieval is efficient and practical.