Fig. 3. Using a PMMC array as a photonic computing core for convolutional image processing.

a Schematic of optical convolution for image processing. An array of k² PMMC is programmed to store the kernel matrix. A patch of pixels of an image is encoded as optical pulses and input into k² optical channels to perform MAC operation with the kernel. The output in TE₀ and TE₁ are summed incoherently and measured with photodetectors. The activation map is represented by the mode contrast and could be both positive and negative. b Optical microscope image of the photonic core consisting of four PMMCs with four input channels. The TE₀ mode outputs are summed on-chip with Y-junctions whereas TE₁ mode outputs are summed off-chip. Optical control pulses are input using the same set of grating couplers used for the TE₁ mode detection. c The greyscale image of “cameraman” (with permission from its copyright owner Massachusetts Institute of Technology) is used as the input image. d, e Left: the raw image generated by convolution with the kernel matrix for detection of horizontal (d) and vertical (e) edges. Right: the corresponding kernel matrix for edge detection. f Combined image of horizontal and vertical edge detection, highlighting all the sharp edges in the original image.