Learning-Based Lensless Imaging through Optically Thick Scattering Media

A team of researchers recently found that images of macroscopic objects hidden behind an optically thick scattering medium can be successfully reconstructed by using a hybrid neural network
27 June 2019
Learning-based lensless imaging through optically thick scattering media
Original article published in Advanced Photonics, https://doi.org/10.1117/1.AP.1.3.036002

Light beams scatter when they encounter optically thick media such as skin or dense fog. An object hiding behind such media cannot be clearly imaged. Conventional optical imaging techniques employ a kind of "gating" to select the part of light that is the least scattered. The least scattered light becomes extinct exponentially along with the optical thickness of the scattering medium. Even though it can be used to form an image, the contrast is usually low owing to the arrival of significant amounts of scattered light during the gating time.

Several methods have been proposed to use the scattered light for image reconstruction. For example, correlation analysis of the scattered light, known as the memory effect, can be used to reconstruct images of mesoscopic objects. But this method fails when the scattering medium becomes optically thick.

As reported in Advanced Photonics, a team of researchers from Shanghai Institute of Optics and Fine Mechanics (SIOM) recently found that images of macroscopic objects hidden behind an optically thick scattering medium can be successfully reconstructed by using a hybrid neural network.

In their demonstration, the scattering medium was a white polystyrene slab 3 mm thick. A hybrid neural network was trained using a set of 3990 labeled pairs of MNIST images and corresponding speckle patterns; it was tested using a set of MNIST data not in the training set. The reconstructed result is shown in the figure below.

Reconstructed results

The reconstructed results. (a) The speckle patterns (64 × 64 pixels) cropped from the raw acquired scattered pattern (512 × 512 pixels), (b) the reconstructed images by using the proposed HNN, (c) the ground-truth images, and (d) the reconstructed images by using memory effect.

In addition, they also demonstrated that an object image could be retrieved from only a small part (0.1%) of its speckle pattern captured by the camera. For instance, in their experiment, they used a camera with 512 × 512 pixels and 16-bit depth to acquire the speckle, but only a small block of 64 × 64 pixels was sufficient to retrieve an object image, even when the speckle was compressed to be binary.

Learning-based lensless imaging offers a powerful technique for forming images of objects in optically thick scattering media, tackling a long-standing and challenging problem in optics.

Read the original research article in the open-access journal Advanced Photonics. Meng Lyu et al., "Learning-based lensless imaging through optically thick scattering media," Adv. Photonics 1(3), 036002 (2019).

Enjoy this article?
Subscribe to SPIE Journal Newsletters
Read the latest research
Recent News