Bioinspired Sensors and Algorithms for Clear Underwater Vision
Abstract
Clear underwater vision in both near-shore and open-ocean waters is of paramount importance for many naval missions involving scuba divers or unmanned undersea vehicles. Optical imaging for clear underwater vision is advantageous over ultrasound-based imaging due to its higher resolution, lower power consumption, compact form factor, and passive sensing properties for clandestine missions. Furthermore, multispectral and polarization imaging techniques capture unique properties of the reflected light from the imaged environment, such as material properties, surface roughness, and 3D shape of underwater objects, which can be used for real-time target identification and classification. However, since visible light has higher scattering and absorption in water compared to ultrasound, optical imaging techniques suffer from lower contrast and lower sighting distances. Although ultrasound imaging can circumvent these problems, it is not well suited to nearshore missions, where high-acoustic-noise environments limit its precision to detect and classify targets.Based on our previous work on underwater geolocalization, polarization image sensors, and polarization signal processing algorithms, we propose to improve underwater vision and target classification by developing a novel bioinspired, multispectral-polarization image sensor and physics-based algorithms augmented with machine learning. We will accomplish these goals by taking an approach radically different from current traditional computer vision approaches and state-of-the-art imaging technology: functionally mimicking the visual system of the mantis shrimpconsidered the best predator in shallow watersto develop a single-chip, low-power, low-noise, high-resolution, multispectral-polarization imaging system. The proposed bioinspired imager will capture 12 spectral bands in the visible spectrum and 3 polarization states, co-registered in real time and with high resolution by monolithically integrating pixelated nanowire polarization filters and dielectric stacks of multispectral filters with an array of vertically stacked photodiodes. The readout circuitry will be optimized for low-noise performance and will enable detection and classification of targets camouflaged by the underwater background. This new imaging sensor will require development of specialized signal processing algorithms, such as interpolation and spectral reconstruction to be able to accurately detect underwater targets.Furthermore, we will develop dehazing or clear water vision algorithms tailored for the proposed image sensor. These algorithms will estimate the background of the imaged scene based on the polarization signatures recorded by our sensor and the geolocation of the underwater camera. We will utilize our physics-based models of the underwater polarization signatures to help distinguish and segment targets from the background. The multispectral information captured by our sensor will then be used to classify underwater imaged objects. We will image a known underwater scene full of potential targets of interest from dusk until dawn over several days. This will enable us to generate a large data set of underwater scenes under various conditions and employ machine learning algorithms to help us determine the best approaches to detect potential underwater hazards.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 06, 2021
- Source ID
- N000142112177
Entities
People
- Viktor Gruev
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign