Scalable imaging through scattering media via physics-informed sparse optronic convolutional neural networks with knowledge distillation

Abstract

Imaging through scattering media is a significant challenge in optical research, and numerous approaches based on deep-learning (DL) have shown impressive progress. Recently, the optronic fully convolutional neural network (OP-FCNN) has demonstrated the potential of opto-electronic DL methods in this domain. However, OP-FCNN suffers from limited generalization due to its reliance on speckle-target training data and the absence of physical priors. Additionally, its deep structure and large number of parameters hinder practical implementation on optical platforms. In this paper, we propose a physics-informed sparse optronic convolutional neural network (PI-OPCNN) that integrates speckle-correlation physical theory with optical neural networks to enhance scalability and reduce network parameters through channel pruning and knowledge distillation. The method incorporates optical lens systems and spatial light modulators (SLMs) to perform speckle autocorrelation as the pre-processing step. Based on our previous work, we designed a U-type optronic convolutional neural network for target reconstruction and applied channel pruning to halve the number of convolutional channels. After fine-tuning with knowledge distillation, the pruned model achieves comparable reconstruction performance to the original while demonstrating high generalization across different diffusers and imaging positions. Experiments involving three diffusers and thirty imaging positions validate the scalability and lightweight characteristics of the proposed approach, highlighting its potential for designing efficient and scalable optical imaging systems for imaging through scattering media.