Shape recognition is necessary in a broad band of applications such as traffic sign or work piece recognition. It requires not only neighborhood processing of the input image pixels but global interconnection of them. The Hough transform (HT) performs such a global operation and it is well suited in the preprocessing stage of a shape recognition system. Translation invariant features can be easily calculated form the Hough domain. We have implemented on the computer a neural network shape recognition system which contains a HT, a feature extraction, and a classification layer. The advantage of this approach is that the total system can be optimized with well-known learning techniques and that it can explore the parallelism of the algorithms. However, the HT is a time consuming operation. Parallel, optical processing is therefore advantageous. Several systems have been proposed, based on space multiplexing with arrays of holograms and CGH's or time multiplexing with acousto-optic processors or by image rotation with incoherent and coherent astigmatic optical processors. We took up the last mentioned approach because 2D array detectors are read out line by line, so a 2D detector can achieve the same speed and is easier to implement. Coherent processing can allow the implementation of tilers in the frequency domain. Features based on wedge/ring, Gabor, or wavelet filters have been proven to show good discrimination capabilities for texture and shape recognition. The astigmatic lens system which is derived form the mathematical formulation of the HT is long and contains a non-standard, astigmatic element. By methods of lens transformation s for coherent applications we map the original design to a shorter lens with a smaller number of well separated standard elements and with the same coherent system response. The final lens design still contains the frequency plane for filtering and ray-tracing shows diffraction limited performance. Image rotation can be done optically by a rotating prism. We realize it on a fast FLC- SLM of our lab as input device. The filters can be implemented on the same type of SLM with 128 by 128 square pixels of size, resulting in a total length of the lens of less than 50cm.
[1]
K Mori,et al.
Optical Hough-transform processor with a two-dimensional array of computer-generated holograms.
,
1994,
Applied optics.
[2]
D Casasent,et al.
High-speed acousto-optic mapping modulator for the generalized Hough transform.
,
1993,
Applied optics.
[3]
G Eichmann,et al.
Coherent optical production of the Hough transform.
,
1983,
Applied optics.
[4]
I Glaser,et al.
Lenslet array processors.
,
1982,
Applied optics.
[5]
Gene Gindi,et al.
OPTICAL FEATURE EXTRACTION VIA THE RADON TRANSFORM.
,
1984
.