Speckle-Based Optronic Transformer for Target Recognition in Scattering Imaging

Light scattering constitutes a prevalent phenomenon across multiple domains. Scattering media including atmospheric particulates like clouds and fog significantly degrade remote sensing imaging capabilities, posing substantial challenges to automated target recognition (ATR) systems. Recent advancements have widely adopted deep learning (DL) architectures for ATR applications leveraging scattered light speckle patterns. However, the escalating computational demands and energy requirements present critical constraints in DL implementation. This study introduces a speckle-based optronic transformer (S-OPT) architecture, implementing the neural network through optical computing paradigms for SAR ATR applications. The methodology employs random phase noise distortion modeling to simulate light scattering phenomena, coupled with specialized optical imaging hardware for capturing MSTAR target speckle patterns. Speckle pattern datasets are employed to train the S-OPT system’s target discrimination capabilities under scattering conditions. The S-OPT architecture achieves full optical computation implementation, demonstrating threefold advantages: nanosecond-scale processing speed, microwatt-level power consumption, and minimal spatial footprint. Experimental validation using MSTAR benchmark datasets confirms the system’s operational viability and architectural efficiency for SAR ATR in scattering environments.