Computational imaging and automated identification for aqueous environments

Abstract

Sampling the vast volumes of the ocean requires tools capable of observing from a distance while retaining detail necessary for biology and ecology, ideal for optical methods. Algorithms that work with existing SeaBED AUV imagery are developed, including habitat classification with bag-of-words models and multi-stage boosting for rock sh detection. Methods for extracting images of sh from videos of long-line operations are demonstrated. A prototype digital holographic imaging device is designed and tested for quantitative in situ microscale imaging. Theory to support the device is developed, including particle noise and the effects of motion. A Wigner-domain model provides optimal settings and optical limits for spherical and planar holographic references. Algorithms to extract the information from real-world digital holograms are created. Focus metrics are discussed, including a novel focus detector using local Zernike moments. Two methods for estimating lateral positions of objects in holograms without reconstruction are presented by extending a summation kernel to spherical references and using a local frequency signature from a Riesz transform. A new metric for quickly estimating object depths without reconstruction is proposed and tested. An example application, quantifying oil droplet size distributions in an underwater plume, demonstrates the efficacy of the prototype and algorithms.