A Neural Net for Reconstructing 3-D Images of Object Surfaces

Abstract

There are some algorithms for reconstructing a 3-D surface from two 2-D images. Usually these algorithms are different for shaded objects and for 3-D surfaces from random-dot stereograms. We here propose a neural net for the reconstruction of images of such objects. Two arrays (matrices) of photosensitive elements act as inputs to the neural net. The real 3-D object (shaded sphere and planes) illuminated by a point light source produces two images in the plane of these matrices. The information from photomatrices is processed by many independent channels. Each channel obtains information only from a delimited area of the matrices. This area is called the receptive field (RF) of the channel. RFs of the different channels overlap. If RFs have the same coordinates on the two matrices, they form a pair of corresponding RFs. The information obtained from this pair is integrally processed, determining both the spatial position of the single object surface fragment and its averaged brightness. Two subsystems (coarse and fine) are used to determine the spatial position. First, the orientations of the matrices are chosen so as to bring down to a minimum the integral differences of light intensities between the images in all corresponding RFs. The coarse subsystems determine the disparities of the corresponding RFs, taking into account the orientation (fixation disparity) of the matrices. The fine subsystems calculate the centroids of the images within the RFs and the averaged intensity across the RFs. This information is used to calculate the exact position of the fragment in space and the averaged brightness. All calculations are carried out by neuron detectors sensitive to object location in space. These neurons form the output layer of the neural net.