Abstract
Face identification relies on the processing of a specific range of spatial frequencies (SF; peaks between 7 and 16 cycles per face [cpf]; e.g. Näsänen, 1999; Willenbockel et al., 2010) and orientations (SO; centered on the horizontal 90 degrees axis in Fourier space; Duncan et al., 2019). However, because SF and SO tunings are typically measured separately, how they are combined to support face identification remains unclear. Previous work compared face identification performance using limited combinations of SO (two bands: horizontal and vertical) and SF (three bands: 4, 16 and 64 cycles per image [2.66, 10,66 and 42,66 cpf]; Goffaux et al., 2011). In the present study, we favored a data-driven method similar to Bubbles (Gosselin & Schyns, 2001), allowing continuous sampling of SF and SO combinations across the whole spectrum. Seven participants performed 1,500 trials on a delayed same/different face matching task to reveal their combined SF-SO tuning. A broadband target face was shown for 300ms, followed by a 500ms white noise mask. A probe face, in which SF and SO were randomly sampled, was then presented until an answer was given. The task was to decide whether the probe matched the target or not. Results reveal a single channel with SF peaking at 15 cpf (width of 2.3 octaves) and SO peaking at 91 degrees (width of 15 degrees) was used to complete the task. This pattern of results replicates prior findings showing that face identification relies on limited information consisting of mid-to-high horizontal frequencies (Goffaux et al., 2011). However, use of a random sampling method gives a more complete picture, accounting for every possible SF-SO combination. Future work could benefit from such an approach to explore important issues like individual or cultural differences, thus potentially paving the way for adapted face training paradigms.