Data-driven classification in optical correlation systems

Optical correlation is a powerful paradigm for object recognition owing to its inherent parallelism and high-speed processing. Nevertheless, the decision-making process based on correlation planes remains limited, as most optical correlators are primarily reliable for binary classification, which confirms only the presence or absence of the target by observing correlation peaks. This limitation becomes particularly critical when query signals contain multiple targets, as conventional approaches require extensive correlation analysis with multiple references, rendering the process complicated and less scalable. To address this challenge, we propose a data-driven framework that operates directly on the correlation plane and effectively classifies query signals by leveraging the physical advantage of the optical correlator and the discriminative power of the neural network. To the best of our knowledge, this unique learning-based correlation approach extends the applicability of optical correlators beyond binary classification. Importantly, training the network directly on correlation maps can reduce dimensionality relative to input images while preserving discriminative features. In this approach, correlation maps are generated using a VanderLugt correlator and then analyzed with two proposed neural network models, enabling robust classification. The proposed method has been investigated using the digit-MNIST, fashion-MNIST, and a laboratory-prepared vehicle dataset, and evaluated using standard figures of merit, confirming high accuracy and precision. This optical-neural approach extends the applicability of optical correlators to a range of real-world applications.