Optical convolution accelerator based on SLM's imaging-casting configuration

Convolutional neural networks (CNNs) take the most computational load of deep-learning multilayered neural networks. This paper presents an optical convolution accelerator based on a spatial light modulator (SLM) imaging-casting configuration, which enables optical parallel computation with high speed. This design employs a 4 f optical system composed of two SLMs and two Fourier lenses, with the SLMs at the object plane and image plane of the system, respectively. This approach effectively overcomes the constraints imposed by the Fourier transform in traditional designs, thereby rendering SLMs more compatible with optical systems and enabling the expansion of computational scale. Furthermore, in comparison with other existing approaches, this scheme theoretically achieves higher computational accuracy for convolution operations and a simpler system structure. Experimental results show that the average relative errors of convolutions are less than 3.50%. This optical convolution accelerator is expected to provide significant advantages towards large-scale optical implementations of CNN framework for the eventual applications of parallel optical computing in the future.