Medical Image Retrieval Based On Feature Fusion

The content-based image retrieval(CBIR) originated in the eighties of the last century, people mainly mark keywords by hands to retrieval image. With the rapid development of the image processing technology, CBIR technology has been widely applied in the field of image processing. The CBIR plays a significant role in the diagnosis and medical research on medical clinical assistant, and it has been got more and more attention. This paper studies a medical image retrieval method based on multi-feature fusion.The paper elaborated the significance of CBIR research, application status, basic principles and key technologies. This paper retrieves by the variety of extracted image features (SIFT, SURF and topological features). Meanwhile, for the insufficient of similarity measure of the SIFT algorithm, we propose a new method with Hellinger kernel function to measure the similarity-ROOTSIFT. On this basis, we can combine the topological characteristics with image features to implement medical image retrieval.For medical images, we use SIFT, ROOTSIFT, SURF and topological characteristics alone to retrieved at first, and then give corresponding weights for each feature based on the good or bad retrieval results. Finally, based on the previous retrieval algorithm, a new method by integrating a variety of low-level visual features is proposed to improve the final retrieval accuracy rate. In the first step, image feature points and the corresponding eigenvectors are gotten by the SIFT (SURF, ROOTSIFT) algorithm. In the next step, to increase local feature similarity, topological features of the SIFT (SURF, ROOTSIFT) characteristics match points are extracted and measured. Finally, points similar both in topological features and SIFT (SURF, ROOTSIFT) characteristics are adopted to be weighted linearly and obtain the most similar image with a reference image. Experimental results demonstrate that the retrieval accuracy by fusing multi-features is higher than by a single feature. Compared with other algorithms, the proposed algorithm can really improve the retrieval accuracy without the loss of true retrieval results, while excluding the retrieval error.