Hdmr method on image retrieval from large-scale databases

This thesis study aims to apply the basic idea of dimension reduction of the HDMR method as a solution to image retrieval from a large-scale database problem. Previous researches on applying the HDMR method on image processing studies showed that this method is succesful in representation of images. During the performed experimental studies the pattern image and all images in the databese work on were subjected to HDMR method to get HDMR terms of images. Then by calculating and comparing the angles between univariate terms of the pattern image and all images in the database the target image was tried to find. The algorithm was setted to retrieval the image of the smallest angle with the pattern image. The HDMR method was used for two purpose in this study. First aim is to retrieve the same images from large scale databases by using a pattern image to determine if the pattern image exists in databases. It is a exact matching. The other purpose is retrieval of the similar images with the pattern image. During the experiments the common databeses for image processing researches of COIL, COREL and LFW were used to test the performance of the HDMR method in image retrieval. For comparison purpose, PCA which is one of the leading dimension reduction methods was used. Since PCA method works with very large matrices on large scale databases, it could not be executed on the databases mentioned above. To overcome this problem, new sub databases are setted from the main databases and for fair comparison the HDMR method was also applied on this new databases. At the end of the experiments it came to sight that the HDMR method is 100 % successful in accuracy for retrievaling the same image with the pattern image (exact matching) for all images from all databases mentioned above. In spite of this, the PCA method reached its highest accuracy with k=0.9 as 100 % in LFW face database, but its accuracy results for COIL and COREL databases are 86.67 % and 86.36 % respectively. Moreover the propopsed method is faster than the PCA method as 3-5 times in terms of the avarage image retrieval time per image. PCA method retrieves images from COIL, COREL and LFW sub databases averagely in 13.83, 21.37 and 34.12 miliseconds respectively when k=0.9 in the case it reaches its highest retrieval accuracy. On the other hand, HDMR method retrieves images from the same databases in 4.33, 5.14 and 6.65 miliseconds respectively. HDMR method reached also remarkable results in the problem of retriving the similar images with the pattern from large-scale databases. This study points up the fact that the HDMR method is a very efficient method in image retrieval from large-scale databases. For advanced researches it can be applied to classificaiton and face recognition problems.