Artificial intelligence (AI) processing for enhancing an intelligent sensor - I

Abstract

Retrieving images from visually random optical speckles poses a coveted yet formidable task across diverse scenarios. In recent years, the advent of Deep Learning (DL)-based approaches has ushered in significant progress, showcasing impressive performance. Nevertheless, prevailing solutions have predominantly relied on singular network architectures to capture the inverse scattering process, yielding less than optimal recovery outcomes. In our study, we introduce a pioneering objective function that incorporates implicit cyclic adversarial loss. We propose a symmetric forward-inverse reinforcement (SFIR) framework, aimed at increasing image recovery performance through scattering media. This innovative framework employs two distinct networks: one dedicated to modeling inverse scattering, and the other to forward scattering processes. Leveraging a symmetric training strategy in alignment with our formulated objective, we fully exploit the feature extraction capabilities and fine-tune network parameters, thereby achieving superior image recovery compared to conventional single-network paradigms. Our extensive experimentation across diverse datasets unequivocally validates the efficacy and supremacy of our proposed framework. Furthermore, our framework demonstrates resilience to varying noise levels, dataset sizes, and network parameters, exemplifying its adeptness in restoring targets from noisy speckles. Notably, our framework showcases promising results, accentuating its potential in enhancing image restoration and recognition performance across domains such as biomedical imaging, holographic display, and optical encryption applications. In abstract, our framework offers significant advancements for enhancing the capabilities of intelligent sensors in image recovery tasks amidst challenging optical environments.