A novel approach for nose-tip detection on 3D face images across pose

AbstractThis paper investigates a novel technique for extraction of the nose-tip from three-dimensional face image in any pose, which is needed for nose-tip based face registration. The present technique uses weighted median filter for smoothing. No normalization process is applied and the system correctly detects nose-tips across any pose variations. In our system, at first the range images are thresholded using Otsu’s thresholding algorithm, then filling of holes is done using interpolation method and after that smoothing is done using weighted median filtering mechanism. In the last and final step, nose-tip is detected using maximum intensity algorithm. To evaluate the performance of our approach for nose-tip localization, we have used FRAV3D, GavaDB and Bosphorus database. In case of FRAV3D database out of 542 range images, nose-tips were correctly located for 536 images thus giving 98.70% of good nose-tip localization, in contrast to the method without smoothing which accounted for only 521 face images. In case of GavaDB database, nose-tip was correctly recognized for 421 images out of 549 images thus giving 76.68% of good nose-tip localization in contrast to the method without smoothing which accounted for only 405 face images. In case of Bosphorus database the recognition rate for nose-tips was far better than FRAV3D and GAVADB because the present technique detects nose-tips correctly for 4476 images out of 4935 correctly with smoothing in contrast to the method without smoothing which accounted for only 4333 face images. The overall performance of the system is 90.27% with smoothing whereas the original system gave a performance of only 87.27%.From the results we can conclude that maximum intensity technique has great capabilities for nose-tip detection across variant poses and that Bosphorus database has given a better result than FRAV3D and GAVADB database. KeywordsFace Recognition, threshold, pose, smoothing, weighted median