Primitive features by steering, quadrature, and scale

The impulse response of neurons in the visual cortex of the mammalian brain has been known for some time. How to make use of these as filters has led to many hypotheses. The response of a single filter is ambiguous because the result depends on stimulus type, contrast, position, orientation, and scale. We show that a set of quadrature filters at sparse positions can be constructed so that it is possible to disambiguate the 2D responses of the individual filters. Detecting edges is not the goal of the present work; rather, we seek to detect relevant edges. Thus, we make the assumption that at the scale of interest, a local image patch consists predominantly of an edge or a bar. When this patch is processed by five or seven oriented filters, one can compute the exact orientation and centroid position of the feature. When the set of filters is applied at two different scales, it is possible to distinguish edges from ridges and to identify the polarity, intensity, and width. It is also possible to find corners and blobs. These computations are stable under image shifts in position and orientation and can be made to subpixel resolution.