Multi-View Acoustic Sizing and Classification of Individual Fish (original) (raw)

Acoustical Imaging, 2011

Abstract

ABSTRACT Estimating biophysical parameters of fish populations in situ such as size, orientation, shape, and taxa is a fundamental goal in oceanography. Towards this end, acoustics is a natural choice due to its rapid, non-invasive capabilities. Here, multi-view methods are explored for classification, size and orientation estimation, and 2D image reconstruction for individual fish. Size- and shape-based classification using multi-view data is shown to be accurate (~10% error) using kernel methods and discriminant analysis. For species-based classification in the absence of significant differences in size or shape, multi-view methods offer significant (~40%) reduction in error, but absolute error rates remain high (~20%) due to the lack of discriminant information in acoustic scatter. Length and orientation estimation are investigated using a parameter-based approach with a simple ellipsoidal scattering model. Good accuracy is obtained when the views span the full 360°. When the span is limited to less than 60°, incorporating a prior constraint on possible body shapes can lead to reduced uncertainty in the estimated parameters. Finally, using views that span the full 360°, sparse Bayesian learning coupled with a conventional Radon transform yields accurate two-dimensional, projected images of the fish. KeywordsAcoustics- Fish classification -Scattering-Size estimation- Radon transform - Bayesian learning -Distorted wave Born approximation

Jules Jaffe hasn't uploaded this paper.

Let Jules know you want this paper to be uploaded.

Ask for this paper to be uploaded.