Purchase this article with an account.
Hayaki Banno, Jun Saiki; Higher-order image statistics is a cue for animal detection. Journal of Vision 2011;11(11):837. doi: 10.1167/11.11.837.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
We can rapidly judge whether images of natural scenes contain animals or not (Thorpe, Fize and Marlot, 1996). So what information do we utilize to rapid animal detection? Image statistics could be a candidate for the task. For example, Torralba & Oliva (2003) found that Fourier amplitude spectra of images potentially predicted the presence/abasence of animals. Nevertheless, for humans, it remains unclear which statistics are important to detect animals. Behavioral studies suggested that human observers are unlikely to utilize amplitude information on its own (Wichmann et al., 2010). We investigated whether higher-order image statistics proposed by the texture synthesis algorithm of Portilla and Simoncelli (2000) could be a cue for animal detection. These statistics are likely to be essential for human texture perception under brief viewing (Balas, 2006), so we hypothesized that the statistics also would contribute to animal detection. In Experiment 1, we compared detection performance between three types of distracter images: synthesized textures which shared statistics with animal images of natural scenes, synthesized textures having the same statistics as non-animal images and non-animal images themselves. Participants had to detect animals in pictures with brief presentation (40 ms) as quickly and as accurately as possible in the Yes/No paradigm. On each trial, a single image was located on the fixation point (0 deg) or at the eccentricity of 14 deg. The detection performance was significantly lower when texture images sharing the statistics with animal images were used as distracter than other distracter conditions at both eccentricities. In Experiment 2, we carried out the same task using new images with their amplitude equalized. The result ruled out the possibility that amplitude difference between animal and non-animal image sets cause the performance impairment. These findings suggest that humans make use of higher-order statistics for rapid animal detection.
This PDF is available to Subscribers Only