The effect of bandpass filtering on the efficiency of complex pattern identification

Purpose. To measure the efficiency with which observers use spatial frequency information when identifying complex visual patterns. Methods. Stimuli were 1 and 2 octave wide bandpass filtered English letters and familiar faces embedded in 2D dynamic Gaussian noise. Center frequencies of the filtered images ranged from 1.4 cycles/image to 90.5 cycles/image. Human and cross-correlator contrast variance thresholds for 67% correct identification were measured for each combination of frequency range and stimulus type in a 10-alternative identification task. Efficiency (defined as the ratio of cross-correlator to human threshold) was calculated in each condition. Results. Unlike previous results, letter identification efficiency was relatively constant across spatial frequency ranges for both 1 and 2 octave wide bandpass filtered stimuli. Surprisingly, observers were unable to identify faces filtered by most 1-octave wide bands. In the 2-octave condition, face identification efficiency was highest at medium frequencies (4-16 cycles/image height) and declined at higher frequencies. Face identification was impossible for the lowest 2-octave frequency range. A subsequent experiment showed that the different results obtained with letters and faces were not due to differences in their respective power spectra. Conclusions. Results suggest there is not one common algorithm used for identification of all complex visual patterns. Further, the relative efficiency wilh which observers use the output of frequency channels depends upon stimulus characteristics other those of the power spectrum.