Soumika Munshi

Signed-negabinary-arithmetic-based optical computing by use of a single liquid-crystal-display panel

Based on the negabinary number representation, parallel one-step arithmetic operations (that is, addition and subtraction), logical operations, and matrix-vector multiplication on data have been optically implemented, by use of a two-dimensional spatial-encoding technique. For addition and subtraction, one of the operands in decimal form is converted into the unsigned negabinary form, whereas the other decimal number is represented in the signed negabinary form. The result of operation is obtained in the mixed negabinary form and is converted back into decimal. Matrix-vector multiplication for unsigned negabinary numbers is achieved through the convolution technique. Both of the operands for logical operation are converted to their signed negabinary forms. All operations are implemented by use of a unique optical architecture. The use of a single liquid-crystal-display panel to spatially encode the input data, operational kernels, and decoding masks have simplified the architecture as well as reduced the cost and complexity.

Photonic implementation of Hopfield neural network for associative pattern recognition

An optical matrix-vector multiplier has ben efficiently used for photonic implementation of Hopfield network model, which is used for binary pattern recognition. Training matrices are recorded on electrically addressed spatial light modulator, where each matrix is composed of the same row of each pattern, that the network is being trained with. After training, if an unknown pattern is presented to the network in the form of a vector, the output vector is obtained by the element that has the highest magnitude through a winner- take-all algorithm. Pattern can be recognized even if the input is noisy and distorted.

Morphological-transformation-based technique of edge detection and skeletonization of an image using a single spatial light modulator

A technique of optically detecting the edge and skeleton of an image by defining shift operations for morphological transformation is described. A (2 × 2) source array, which acts as the structuring element of morphological operations, casts four angularly shifted optical projections of the input image. The resulting dilated image, when superimposed with the complementary input image, produces the edge image. For skeletonization, the source array casts four partially overlapped output images of the inverted input image, which is negated, and the resultant image is recorded in a CCD camera. This overlapped eroded image is again eroded and then dilated, producing an opened image. The difference between the eroded and opened image is then computed, resulting in a thinner image. This procedure of obtaining a thinned image is iterated until the difference image becomes zero, maintaining the connectivity conditions. The technique has been optically implemented using a single spatial modulator and has the advantage of single-instruction parallel processing of the image. The techniques have been tested both for binary and grey images.

Improved optical matrix-vector multiplier

An improved optical matrix-vector multiplication is performed by convolution process. The multiplicated binary numbers are represented by on/off states of light sources and the multiplier binary numbers are recorded on a spatial light modulator. Cylindrical optics is used as free space interconnection. The convolution coefficients are recorded on a CCD array. The output of the CCD array are added in a computer to yield the result of multiplication. The operation is completely digital and needs no analog to digital conversion. Because of parallel operation in two dimensions, the processing speed is greatly increased.