Human face recognition through moment descriptors

Selection and implementation of an appropriate feature extraction technique is a crucial factor for attaining high recognition accuracy in any field of image processing. Several Statistical and Structural based methodologies have been proposed for facial feature extraction. This paper focuses on the well known statistical moments-based, Zernike Moments (ZMs) and the Pseudo-Zernike Moments (PZMs), as well as two-dimensional Polar Harmonic Transform (PHT) descriptors [1]. The latest survey of the literature has shown the application of PHTs only in the field of fingerprint recognition [2] and character recognition [1]. This paper experiments their importance for face recognition. The performance is analyzed and compared with the existing ZMs and the PZMs in terms of accuracy, computational complexity, invariance, robustness to noise and reconstruction ability. The accuracy is evaluated through the widely used ORL Database comprising of 400 face images of 40 people with slight variations in pose, expressions, lighting and facial occlusion [3]. To verify its adaptability for rotated pose variations between 0° - 90°, the well known UMIST pose face database comprising of 575 images of 20 individuals has been explored [4]. The overall accuracy is evaluated through the Nearest Neighbor Classifier. High recognition accuracy of 97.5% is obtained for PHTs on the ORL database as compared to 95.3% and 94.8% achieved by PZMs and ZMs respectively. Experimental results also show that PHTs perform better than ZMs and PZMs on scale invariance, rotation invariance and noise invariance achieving 97.25%, 98.7% and 94% accuracy respectively as achieved by [5] and possess very low computational complexity because the time required for computing their radial kernels is considerably less.