A symmetric forward-inverse reinforcement framework for image reconstruction through scattering media

Abstract

Image retrieval from visually random optical speckles is a desired yet challenging task in various scenarios. Deep learning (DL) based approaches have rapidly grown to achieve impressive performance in recent years. However, the majority of solutions thus far have been confined to a single network to model the inverse scattering process, resulting in relatively poor recovery performance. In this paper, we introduce a novel objective function to embed implicit cyclic adversarial loss and propose a symmetric forward-inverse reinforcement framework congruent with this objective function for enhancing image recovery performance through scattering media, where two networks are designed to model inverse and forward scattering processes, respectively. A symmetric training strategy, aligned with the formulated objective function, is utilized to fully exploit the feature extraction ability and fine-tune the parameters of the networks, achieving higher-fidelity image recovery than that by a single neural network. Both simulation and experimental results on various datasets demonstrate the effectiveness and superiority of the proposed framework. Moreover, this framework also shows convincing robustness to varying noise levels, dataset volumes, and network parameters, indicating proficient restoration of targets from noisy speckles and maintaining comparable performance even with limited training data and fewer network parameters. Furthermore, the promising recovery results on real-world hazy datasets demonstrate that the proposed framework could open up new opportunities to enhance image restoration and recognition performance in biomedical imaging, optical encryption, and holographic display.