Imaging reconstruction through strongly scattering media by using convolutional neural networks

Abstract

As an object passes through strongly scattering slab, it cannot be imaged, due to scattering slab scrambling the object information. In this paper, we employ convolutional neural network (CNN) to realize the imaging reconstruction in such a specific optical system where an object is located between two strongly scattering slabs. The influence of the thickness of the scattering slabs on the imaging reconstruction is investigated, in which we can obtain the imaging reconstruction from the object–speckle pairs of the object passing through the certain thickness of scattering slab, by using the trained CNN with the certain object–specklepairs. It is shown that good reconstructed images are achieved for three different thickness of the scattering slab. Moreover the so-called hybrid trained network is employed, in which we train the network by using totally 9000 object–speckle pairs from three sets of data, each set of data we selecting equal 3,000 pairs of object–speckle​ patterns, corresponding to the three different thickness of the rear scattering slabs. We find that hybrid trained network is capable of reconstructing the image from speckles taken from the three different scattering conditions Our approach may have applications in biomedical imaging.