Registration of Face Image Using Modified BRISK Feature Descriptor

Abstract

Automatic face recognition is a hot area of research in the field of computer vision. Even though a lot of research have been done in this field, still researchers are unable to develop an algorithm which can detect the face images under all possible real time conditions. Automatic face recognition algorithms are used in a variety of applications such as surveillance, automatic tagging, and human-robot interaction etc. The main problem faced by researchers working with the above real time problems is the uncertainty about the pose of the detected face, i.e. if the pose of the sensed image differ from the images in the trained database most of the existing algorithms will fail. So researchers suggested and proved that the detection accuracy against pose variation can be improved if we considered image registration as a preprocessing step prior to face recognition. In this work, scale and rotation invariant features have been used for image registration. The important steps in feature based image registration are preprocessing, feature detection, feature matching, transformation estimation, and resampling. In this work, feature detectors and descriptors like SIFT, SURF, FAST, DAISY and BRISK are used. Among all these descriptors the BRISK descriptor performs the best. To avoid mismatches, using some threshold values, a modified BRISK descriptor has been proposed in this work. Modified BRISK descriptor performs best in terms of maximum matching as compared to other state of arts descriptors. The next step is to calculate the transformation model which is capable of transforming the coordinates of sensed image to coordinates of reference image. Some radial basis functions are used in this step to design the proper transformation function. In resampling step, we used bilinear interpolation to compute some pixels in the output image. A new algorithm is proposed in this work to find out the possible image pairs from the train database corresponds to the input image, for doing image registration. In this work, image registration algorithms are simulated in MATLAB with different detector-descriptor combination and affine transformation matrix. For measuring the similarity between registered output image and the reference image, SSIM index and mutual information is used.