Effects of high-pass and low-pass spatial filtering on face identification

Abstract

If face images are degraded by block averaging, there is a nonlinear decline in recognition accuracy as block size increases, suggesting that identification requires a critical minimum range of object spatial frequencies. The identification of faces was measured with equivalent Fourier low-pass filtering and block averaging preserving the same information and with high-pass transformations. In Experiment 1, accuracy declined and response time increased in a significant nonlinear manner in all cases as the spatial-frequency range was reduced. However, it did so at a faster rate for the quantized and high-passed images. A second experiment controlled for the differences in the contrast of the high-pass faces and found a reduced but significant and nonlinear decline in performance as the spatial-frequency range was reduced. These data suggest that face identification is preferentially supported by a band of spatial frequencies of approximately 8-16 cycles per face; contrast or line-based explanations were found to be inadequate. The data are discussed in terms of current models of face identification.The questions of whether the information concerning the identity offaces is carried by a limited range ofspatial scales and whether the potential information from different regions ofthe spatial spectrum is given equal weight in the determination of identity have been approached in a number of different ways. One method ofconsidering these issues has been to make use ofspatial-frequencyfilteringtechniques (Harmon, 1973). However, variations in this method have produced contradictory results, with notably differentconclusions about the relative importance of different spatial-frequency bands specified in terms ofcycles per face. The term cycles perface is defmed as the number of sinusoidal repetitions of a given width that can be placed within the eye-level width of the face. The use ofthis metric to describe the information present in stimuli allows discussion ofthe degree of detail necessary for recognition, perhaps by defining the scale of facial configuration. A class ofobjects has a configuration if there is a consistent set of features all arranged in the same order. Thus, if a set ofexamples are superimposed, normalizing for scale and viewpoint, another example of the class is produced that is closer to the prototype. Clearly, faces have this property, since all have two eyes, a nose, and a mouth-and these are consistently arranged.Harmon ( can be seen in Figure 2. The images are formed by placing a regular square grid across the image and setting the pixel value at each grid square to the average gray level within it. This work suggested that the minimum image quality that allows effective identification corresponds to a 16 X 16 pixel image; however, since the images did not take up the whole of the screen, the number of pixels per face was slightly lower. Harmon also used a smooth low-pass filtering technique. This type of filtering operation does not introduce additional spatial frequencies (noise), as the pix...

Organization of face and object recognition in modular neural network models

“…Costen et al (1996) have investigated the effect of high-pass and low-pass filtering on face images in isolation, and Parker, Lishman and Huges (1996) have investigated the effect of high-pass and low-pass filtering of face and object images used as 100 ms cues for a same/different task. Their results indicate that relevant high-pass filtered images cue object processing better than low-pass filtered images, but the two types of filtering cue face processing equally well.…”

“…Costen, Parker and Craw (1996) showed that although both high-pass and low-pass image filtering decrease face recognition accuracy, high-pass filtering degrades identification accuracy more quickly than low-pass filtering. Also, Schyns and Oliva (1999) have shown that learning the identity of a set of faces later biases subjects' perception toward low spatial frequency information.…”

Distinct information critically distinguishes judgments of face familiarity and identity.

“…They have also confirmed the importance of the midband of spatial frequencies (8-25 cycles per face), which may provide the optimal information for judgments of face identity (see Costen, Parker, & Craw, 1994, 1996Nasanen, 1999;Ruiz-Soler & Beltran, 2006). Crucially, however, no reverse correlation study (to the best of our knowledge) has explored the visual information driving familiarity decisions (i.e., without an explicit naming component).…”

Face Recognition Is Affected by Similarity in Spatial Frequency Range to a Greater Degree Than Within-Category Object Recognition.

“…Bachmann (1991), using a delayed match-to-sample task, found a rapid decrease in accuracy as coarse-quantized face images were reduced in resolution from 18 pixels per face width to 15. In related experiments, Costen et al (1994Costen et al ( , 1996, using a similar method to Bachmann's, found abrupt drop-offs in performance as faces were low-pass filtered at cutoffs less than 8 cycles per face or high-pass filtered at greater than 16 cycles per face. In experiments comparing human delayed match-to-sample performance with that of an ideal observer algorithm, Nasanen (1999) found a reduction in the efficiency function of participants when frequencies around 8 -13 cycles per face were masked by noise.…”

“…The researchers suggested that this lack of a critical band effect in their data was based on the contributions of overlapping and interfering frequencies. They pointed out that studies such as those of Costen et al (1994Costen et al ( , 1996 used full-bandwidth pictures at learning, meaning that there was a large range of frequencies in the learned image that were not present in the tested image. These frequency elements may have interfered with recognition, as they would have found no equivalents in the tested images.…”