Linearity across spatial frequency in object recognition

In three experiments, we measured recognition as a function of exposure duration for three kinds of images of common objects: component images containing mainly low-spatial-frequency information, components containing mainly high-spatial-frequency information, and compound images created by summing the components. Our data were well fit by a model with a linear first stage in which the sums of the responses to the component images equalled the responses to the compound images. Our data were less well fit by a model in which the component responses combined by probability summation. These results support linear filter accounts of complex pattern recognition.