Face Recognition using SIFT by varying Distance Calculation Matching Method

Abstract

Scale Invariant Feature Transform (SIFT) is a method for extracting distinctive invariant feature from images [1]. SIFT has been applied to many problems such as face recognition and object recognition [18], [19], [20], [21]. We have analyzed performance of SIFT using Euclidean distance as a matching algorithm. Further the matching rate can be enhanced/improved by changing distance calculation methods used for matching between two face images. So this paper also describes face recognition under various distance calculation methods like Correlation and Cosine. The experiments are conducted on different images of ORL face database [17] and Indian Face database [16] by changing illumination condition, scaling and rotation. From the experiments, it is shown that cosine and correlation distance calculation methods have performed well compared to the Euclidean distance matching method of original SIFT.

Integration of optimal spatial distributed tie-points in RANSAC-based image registration

“…The objectives of the experiments were: (1) to provide the suggested empirical values of the parameters for the proposed matching algorithm, and (2) to make a comparison on the effectiveness of the standard RANSAC, the bucketing-RANSAC and the proposed SRS-RANSAC. To obtain the same initial matches, Lowe's (2004) original implementation of SIFT is used for each image pair. In order to evaluate the performances of three methods, we have included comparison of matching accuracy from obtained matches and registration accuracy in the experiments.…”

“…Dare and Dowman (2001) proposed an improved model based on the combination of multiple feature extraction and feature matching algorithms. SIFT operator is a processing chain for feature detection, which was introduced in 1999 and improved in 2004 by Lowe (2004). It is capable of detecting scale and affine invariant features automatically.…”

Hybrid classifier using SIFT descriptor

“…In fact, it consists to find in the second image J the best matching of each founded IP in the first image I using Nearest Neighbor (NN). NN is defined as the IP that has the minimum Euclidean distance (between two descriptor vectors) given by eq.6 [2], every IP is represented as a 128 dimensional vector.…”

“…IP's identification is frequently used in many tasks related to computer vision field such as: object recognition [1], face detection [2], tracking [3], stereo-vision ...). IP's detection is widely employed in image matching.…”

“…It is a robust technique for feature extraction [2] and reliable method for various objects recognitions [10]. It is a very stable descriptor and moderately invariant to changes in rotation, illumination, scaling [11].…”