Imaging through scattering media

Abstract

This project addresses the challenge of imaging through scattering media, a common issue affecting various fields, from autonomous driving through fog to health diagnostics through tissue imaging. Conventional approaches, such as deconvolution and speckle correlation, have their own limitations such as the invasive nature and time-consuming computation. Recent developed approach with machine learning (ML) offers a promising non-invasive option. It can extract important system features without knowing the detailed knowledge of the physical principles. Our work integrates machine learning with conventional methods to reconstruct scattered images, focusing on varying speckle sizes. The study is purely software-based, simulating the imaging and reconstruction process to validate the proposed methodology. We utilize a U-net model, a form of a convolutional neural network known for its effectiveness in image segmentation and denoising, to predict images from unseen point spread functions (PSFs), which demonstrates the model's ability to generalize across different scattering conditions. The results show the U-net model's potential in reconstructing images through unseen scattering media, making a significant step forward in non-invasive imaging techniques. The findings in this study can be used to accelerate the development of better ML models to reconstruct scattered images in the field of image processing.