Sandwich GAN: Image Reconstruction from Phase Mask based Anti-dazzle Imaging

Conventional camera systems are susceptible to the adverse effects of laser dazzle, which may over-saturate an image or cause permanent damage to pixels. To address this problem, we developed an approach combining point spread function engineering whereby a wavefront-coded mask in the pupil plane blurs both the laser and scene, together with a deep neural sandwich network. In addition to protecting the sensor, our approach jointly removes the laser from the scene and reconstructs a satisfactory deblurred image. Image recovery is achieved by wrapping two generative adversarial networks (GANs) around a learnable non-blind image deconvolution module. We trained the Sandwich GAN (SGAN) to suppress the peak laser irradiance as high as 10⁶ times the sensor saturation threshold - the point at which the bare system without the phase mask may exhibit damage. The end-to-end training includes physics-based modeling of the imaging system whereby a laser having an arbitrary angle of incidence is superimposed on images from a large publicly available library. The trained system was validated in the laboratory for laser strengths up to 10⁴ times the saturation value. The proposed image restoration model quantitatively and qualitatively outperforms other methods for a wide range of scene contents, illumination conditions, laser strengths, and noise characteristics.