Accuracy Improvement of Bag-of-Feature Based Color Medical Image Retrieval with Optimal Selection Initial Point in K-means Clustering

Due to the increasingly progress of medical equipment and medical imaging machines, daily, a large number of medical digital images are produced in therapeutic centers and stored in databases. A large number of these images has changed their retrieval to one of the challenges in this field. One of the most common ways of retrieval of images is a retrieval based on a bag of visual words. One the most popular techniques for producing visual words is using K-means clustering. However, the effectiveness of K-means clustering is highly dependent on the initial selecting centroids for clusters, since in the base algorithm, K-means clustering of clusters centroids are randomly selected. Image retrieval techniques which are based on a bag of words using K-means, are not so effective. The goal of this paper is showing a way for the improvement of the accuracy of colored medical images retrieval content based using the optimized selected points in K-means clustering. In this method after the extraction of local features of images using SIFT descriptor, we start producing the visual words based k-means clustering. Three algorithms have been used for the selection of the optimized initial centroids: based on difference range, based on weight and based on mean. By doing experiments on the collection of medical images, use of the initial optimized points can cause the production of the optimized visual words and provide more accuracy of production of visual words with the initial random centroids.

Keywords:

Image retrieval , Bag-of-words , K-means clustering , initial centroids , SIFT descriptor , Optimized visual words