Influence of background on the spatial-frequency channel for object recognition

Campbell & Robson (1968) found that grating detection and discrimination are mediated by spatial-frequency-selective filters. Motivated by this, later work found a single “channel”, always an octave-wide, for the recognition of letters, faces, and natural objects. However, it is unclear what properties of the image influence this recognition channel. Can we break the octave-wide channel result by altering image properties? Recognizing an object requires grouping of the relevant features in the presence of possibly irrelevant background features (Wertheimer, 1923). To separate objects from background in this way, an observer might use several available cues. If spatial frequency is one of these cues, i.e., if the observer extracts the object using spatial-frequency features absent in its background, we would expect that simply removing the background would allow the observer to use more spatial frequencies for recognition. We test this hypothesis by measuring the spatial-frequency channel for a 16-way ImageNet object categorization task separately in the presence and absence of image background. We find that removing the background increases the median channel bandwidth from 1 to roughly 1.5 octaves. Thus, without a background, people can use more frequencies to recognize an object.