Abstract
Psychophysical experiments frequently use random pixel noise to study the coding mechanisms of the visual system. However, interactions in natural scenes occur among elements of larger articulated structures than pixels. We examined how structure in noise influences the ability to recognize objects with a novel coherence paradigm for object recognition. Grayscale images of 200 everyday objects from 40 categories were analyzed with a multi-scale bank of Gabor-wavelet filters whose responses defined the positions, orientations and phases of signal Gabor patches that were used to reconstruct the original image. The proportion of signal to random noise Gabors was varied using a staircase procedure to determine a threshold supporting object recognition on 75% trials. The noise structure was controlled to produce Gabor chains of varying length (1, 2, 3, or 6 elements) and local orientation, forming straight, smoothly or irregularly curved contours. Each trial, nineteen naïve subjects assigned the reconstructed image to one of four categories, randomly selected from all categories. Object recognition thresholds were invariant to the nature of the underlying local orientation structure of the noise. Increasing the length of the noise contours from 1 to 2 elements increased thresholds (p<0.001), but not for longer contours, suggesting that object identification is based on grouping local contour fragments, but only over very local areas. The invariance of performance with respect to feature type suggests that object recognition relies upon prior knowledge of object form, and that such a top-down mechanism allows the visual system to discount variation in local noise. Sensitivity to the number of elements suggests that assembling descriptions that can support the top-down mechanisms depend on the complexity of the distractors.
Supported by NIH R01 EY018196.