Matrix convolutional operation architecture based on optical Fourier transform

In this paper, We propose a highly parallel optical matrix convolution architecture based on optical Fourier transform (OFT). The matrix multiply-accumulate (MAC) operations consists of matrix multiplication and addition. In the matrix multiplication section, the light intensity signals of matrix A and matrix B carried by two SLMs are used. Matrix C is obtained by matrix multiplication by projecting the uniform light intensity of matrix A onto matrix B through a Dammann grating (DG). The matrix C is received by a CMOS camera after a lens, and the spot array of matrix C is summed up, The convolution results of matrix A and B are obtained. This matrix convolution architecture provides an interesting method for large-scale matrix convolution. The optical matrix convolutional architecture has advantages such as high parallelism, high accuracy, and low power consumption. In the future, the optical computing will be perfectly applied in areas such as deep learning algorithms.