Improvement of ear recognition rate using color scale invariant feature transform

• Main Author: Hadidi, Komeil
• Format: Thesis
• Language: English
• Published: 2013
• Online Access: http://psasir.upm.edu.my/id/eprint/47567/1/FK%202013%2040R.pdf; http://psasir.upm.edu.my/id/eprint/47567/

Local features are effective for ear biometrics. Scale Invariant Feature Transform (SIFT) technique has been used in many biometrics types as well as ear, but it is suitable for gray-scale images. In addition, the number of keypoints which can be retrieved by SIFT has an upper limit. This research is aimed to develop a method for using color information (in addition to gray images) to generate additional feature points for higher recognition rate. SIFT has four stages. The first stage of SIFT, which is applying difference of Gaussian function on the image, has been changed such that the resulting key-points will be generated from a pair of RGB color planes. This structure is inspired by color double opponent neuronal circuits in the primate brains. In the last stage of SIFT, the gray and color features will be compared against gray and color database, respectively. The scores of all active color channels will then be added together to produce final score of database images to win as a matching image. The proposed approach is compared with standard model of SIFT by applying both of them on USTB database of ears with 780 side view ear images from several viewpoints up to 20 degrees difference. Comparison among standard and different color opponent channels demonstrates that 4.3% higher recognition rate has been achieved by utilizing Red/Green opponent channel, in addition to the gray channel, for 20 degrees rotation in viewpoint. For Yellow/Blue channel, the improvement is 6% in maximum rotation of the head. Comparative analysis demonstrates that the proposed method can achieve higher recognition rate by utilizing color image information.