Hybrid Optical Diffractive Neural Networks (HODNNs) for Speckle Reconstruction and Physical Auto-encoding

Abstract All-optical Diffractive Deep Neural Network (D2NN) architecture can learn to implement various functions after the deep learning-based design of passive diffractive layers. D2NN offers unique advantages for parallel processing, speed of light processing, and low power consumption. Here, we extend by proposing Hybrid Optical Diffractive Neural Networks (HODNNs) for speckle reconstruction and physical auto-encoding. HODNNs include an all-optical Diffractive Deep Neural Network (D2NN) for light-speed, energy-efficient phase information extractions and electrical convolutional neural networks (CNNs) for high-performance feature extraction with less hardware complexity. The performance of HODNNs for speckle reconstruction and lens imaging is very comparative with NPCC, SSIM, and PSNR reach -0.98, 0.98, and 27dB, respectively. In the future, the presented frameworks can be used for deep tissue imaging and lensless microscopic imaging.