Abstract
Fish morphological feature extraction based on shape alignment is used to estimate the dimensions, detect malformations, locate body parts like the eyes or the gills, classify fish orientation and species, etc. The Ensemble of Regression Trees machine learning approach is employed and specifically, the Deformable Shape Tracking package is adapted for fish shape alignment. Fish farms can benefit from the proposed approach for fish health assessment, fish growth and feeding needs estimation and harvest time selection. It can also be used for fish monitoring in open seas. Eighteen (18) landmarks are used to define the shape of a fish with an accuracy of approximately 95%. The training and testing were conducted using a custom dataset with low quality underwater images displaying seabream fish. The novelty of this approach is that the customized DEST package is implemented on the reconfigurable hardware of an embedded platform to support hardware acceleration and real time operation
