A Comparison of Gabor Filter Methods for Automatic Detection of Facial Landmarks

This paper presents a systematic analysis of Gabor filter banks for detection of facial landmarks (pupils and philtrum). Sensitivity is assessed using the statistic, a non-parametric estimate of sensitivity independent of bias commonly used in the psychophysical literature. We find that current Gabor filter bank systems are overly complex. Performance can be greatly improved by reducing the number of frequency and orientation components in these systems. With a single frequency band, we obtained performances significantly better than those achievable with current systems that use multiple frequency bands. Best performance for pupil detection was obtained with filter banks peaking at 4 iris widths per cycle and 8 orientations. Best performance for philtrum location was achieved with filter banks with 5.5 iris widths per circle and 8 orientations. Gabor filter banks [4] are reasonable models of visual processing in primary visual cortex [5, 3] and are one of the most successful approaches for processing images of the human face [6, 1, 2, 8]. The success of the approach parallels the success of bandpass filter banks, which approximate signal processing in the cochlea, in speech recognition problems. While the optimal filter bank characteristics have been extensively studied in the speech recognition literature, little work has been done to systematically explore which frequency and orientation bands are optimal for face processing applications. The goal of this paper is to start addressing this gap in the literature. To evaluate performance of the different filter bank approaches, we use a standard recognition engine (nearest neighbor) and measure sensitivity using the statistic. This is a non-parametric measure of sensitivity commonly used in the psychophisical literature. 1 Gabor Filters In the space domain, the impulse response of Gabor filters is a Gaussian kernel modulated by a sinusoidal planewave (1) where is a complex sinusoidal, known as the carrier, "!$# % & '( *),+ -(./ 1032,./ (2) The parameters . and 2 . define the spatial frequency of the sinusoidal, in Cartesian coordinates. This spatial frequency can also be expressed in polar coordinates as magnitude 4 . and direction 5 . . The function is a 2-D Gaussianshaped function, known as the envelope 6 7 98:!;#<%= &>+ ?<@ A&B C., @ D70FEG@ &B H.I @DI (3) where . . is the peak of the function, ? , E are scaling parameters, and the J subscript stands for a rotation operation