Shape-based fish recognition via shape space

Abstract

Automatic fish recognition is a recent research work which is needed to assist marine scientists. Among most discriminative features, the fish outline is very efficient for fish recognition. In a previous work, we proposed a method for pattern recognition (classification and retrieval) based on signal registration and shape geodesics. In this paper, we introduce a preliminary step of pose estimation for accelerating the processing time. We then show that shape geodesics may also be used for outline-based fish recognition. Experiments conducted on the SQUID database which is used as a benchmark to evaluate fish shape recognition, show (1) a reduction in computation time of a factor of ten in average, and (2) the outperformance of the proposed scheme compared to previous methods.