Image Retrieval In Dct Domain

Image Retrieval:

Image retrieval in DCT domain Mozammel-Bin-Motalab (Instructor) Institute of Information Technology University of Dhaka BANGLADESH Email: [email protected]

Abstract This is a proposal to image retrieval for a lot of database of any organization dealing with images. The listed image of that organization may be sorted by name or category. If we want to organize the image by its pattern, we need to identify the exact pattern of that image. This work can be done by taking sample of the full image. Even if we want to check the duplication of any individual image we need to use the image retrieval process. Here the most effective and dynamic process named Discrete Cosine Transform (DCT) is used. We then discuss about this algorithm that, given a test image, would search the database for similar images. Our main issue of this proposal to discuss about the 1. Image Retrieval and 2. Image Indexing by DCT domain

The algorithm would return a series of images from the database ranked in order of similarity. This algorithm would be a very useful tool for image retrieval. It is often desirable to have a large sampling of similar images of a certain part of the anatomy for comparison or study. Finding the appropriate images in existing databases can be very cumbersome. To create such database is a big and challenging issue for us. To create a data base algorithm should follow this steps a. As a test set for such algorithms have permission to collect image sets from different sources in the area even from a remote place. b. Over a period of time built up a database of huge images. c. The breakdown by image type for a sample image for each classification in the order listed in some table. d. To reduce the computational intensity and provide initial filtering of detail. The sorting of database is completed. Now the method to categorizing the image will began. This method is introducing the DCT analog to closest match. If the training set is relatively large, it becomes quite likely that we have a very close match to most test images somewhere in the training set. We should then be able to get better classification comparing the test image to each image in the training set. We didn't implement this approach, as we were opting for quicker methods. The largest DCT coefficients are computed for each image in the training set. The N test image DCT coefficients are then

1

compared against the DCT coefficients of each training set image. Classification is based on which training image yields the lowest DCT coefficients. This work extremely well with a large training set, but is computationally much more expensive than Average Image. Retrieval Method This section describes each of the algorithms we implemented and tested for similar image retrieval. Each image retrieval algorithm takes as input a test image and returns a ranking by similarity for each image in the database. Performance results for each of these methods are presented in the results section. The methods are a. Fast Image Querying: This method intends to facilitate rapid shape matching by dispensing with the expensive arithmetic operations that are normally required b. Creating Image Domain: The mean squared error of the test image is computed with respect to the each image in the database table. Then Images are sorted by the category. c. Creating DCT Domain: The DCT of the test image is computed. Then the mean squared error of the largest N coefficients of the test image to the DCT coefficients of each image in the database is computed. At last images are sorted again by category and classification.

These figures describe each of the DCT domain for an image tested for similar image retrieval. Each image retrieval algorithm takes as input a test image and returns a ranking by similarity for each image in the database. Performance results for each of these methods is presented in the results section In order to obtain a quantitative handle on the clusters, we have to examine all of the points in a given classification. From this cloud of points, we examine the variance of the points from the mean position of the cloud. We repeated this to get a variance for each different category. This gave us a measure of how "big" the clouds were. We then looked at the distances between the centers of all of the different clouds. When these "spheres" in N-dimensional space begin to overlap, classification will get worse. When they are well separated, we expect better classification results [ref 1]. Effects of DCT in image retrieval: It is interesting that retrieval performance was generally good if DCT is used. This is actually not surprising that the much wider variance in images. For image retrieval, the DCT methods produced good matches under both the subjective and quantitative metrics. Nonetheless, it would be necessary to have a larger test database to truly

2

compare the running times between our various techniques. Direct implementations in a compiled language would also improve execution time [ref 2]. The classification and retrieval algorithms could be improved by applying importance weightings on the transform coefficients. Often, the proximity of the query image to a database image in one transform basis dimension can be disproportionately important in terms of clustering.

Reference [1]www.thesis.lib.ncu.edu.tw/ETDdb/ETD-searchc/getfile?URN=83345009&filename=83 345009.pdf [2]www.hal.t.u-tokyo.ac.jp/en/research [3]http://spiedl.aip.org/getabs/servlet/Get absServlet?prog=normal&id=OPEGAR0 00045000004047003000001&idtype=cv ips&gifs=yes

Discussion: I found the topics very interesting but comparatively tough to understand. So far I understand, this subject will be very much useful in near future. Most use of this technique is in web based search engine, website for image and photography, database that containing image and last but not the least give special facilities in social website to find a special person.

[4] C. Jacobs, A. Finkelstein, D. Salesin. "Fast Multiresolution Image Querying". Proceedings of SIGGRAPH, 1995.

Recently this image indexing is using to – 1. Medical Image processing and matching similar X-ray, MRI, MR image to make decision about disease. 2. Math DNA profiling system 3. Extract the suspected person from an image etc. Though the Image retrieval in DCT domain is not implemented fully yet, it is assumed to be a fine implementation of this in near future.

3