Processing the output signals of joint transform correlators using a pre-trained convolutional neural network

Subject of study. This study presents coherent opto-digital joint transform correlators (JTCs) with neural-network-based post-processing of output signals. Aim of study. The aim is to experimentally evaluate the feasibility of applying a deep-learning neural network to process the signals of an opto-digital diffractive JTC for image-recognition tasks. Method . A convolutional neural network (CNN) was used to process the output signals of the opto-digital diffractive image correlator. The input images provided to the correlator had a resolution of 256×256pixels, while the fragments of the correlation functions analyzed by the neural network had a resolution of 32×32pixels. This dimensionality reduction facilitates the combination of high-speed optical processing with the flexibility of neural-network-based analysis. Main results. Two variants of a JTC employing modern spatial light modulators were implemented: one based on a digital micromirror device and the other on a liquid crystal modulator. To classify the correlator output signals, a CNN was used. This network was pre-trained on numerically simulated correlation responses generated using invariant correlation filters designed to produce a specific autocorrelation peak shape. These filters were synthesized from image sets that were distinct from those used in the experimental phase. The study demonstrates that, in both implementations, neural-network-based processing of the correlator output enables successful recognition of input images. Practical significance. These findings can be utilized in the development of high-speed image-recognition systems for various applications.