Abstract
For several years we have been studying an adaptable, rectifying, contrast-comparison process (nicknamed “Buffy adaptation”) that dramatically increases visibility for some patterns and decreases it for others. We have been using patterns composed of evenly-spaced Gabor patches. Such patterns seem rather different from continuous texture patterns like contrast-modulated noise. Here we investigate whether the adaptable rectifying contrast-comparison process can also be demonstrated with contrast-modulated noise. And we ask whether a simple model can account quantitatively for the results.
The adapt pattern (1 sec in duration) is binary noise: an array of 512×512 checks in which each check is randomly chosen with probability 0.5 to be one of two luminances. The contrast of the binary noise adapt pattern is 50%. The test pattern (∼90 ms in duration) has stripes defined by binary noise of two different contrasts C1 and C2 (with a square-wave modulation frequency of approximately 0.5 c/deg). The observer's task is to identify the orientation of the contrast-defined stripes, which can be either horizontal or vertical. To measure an observer's threshold, we hold the average test contrast constant and vary the difference between the contrasts (|C2-C1|). When one of the test contrasts is less than the adapt contrast and one is greater, it is called a straddle test pattern. We find the same results as with the Gabor-patch patterns: Performance is substantially worse on straddle test patterns than on test patterns in which both C1 and C2 are on one side of the adapt contrast (as long as they are not too far above or below the adapt contrast, in which case performance is again poor).
A simple model containing the adaptable rectifying contrast-comparison process in conjunction with a conventional gain-control process (e.g. a normalization network) can account well for the results.