This paper presents an experimental computer vision system which automatically assigns identity to digital images of individual white sharks (Carcharodon carcharias). Our approach utilises dorsal fin posterior edge morphology as a machine-recognizable biometric identifier for differentiating individual sharks. The system takes as input high-quality, high-resolution images of a side view of the fin, which can be routinely acquired during boat trips to the species' natural habitats. After applying lighting correction using histogram equalization, dorsal fin containing regions are identified using the Viola-Jones object recognition framework. Subsequently, fin area subregions are oversegmented using a network of shape-feature guided local optimisers, before exact fin outlines are implicitly described by the merger of image segments into semantic fin shapes, based on machine-learnt fin textures. Finally, individual identity is assigned based on the statistical integration of evidence from multiple local and global shape features.
The jagged pattern of the posterior edge of the dorsal fin is individually characteristic and its morphology is stable over decades. Teams of human experts are able to accurately identify individuals from these patterns with a genuine acceptance rate of up to 98%. However, manual identification is highly labour intensive. Here we compare the performance of our algorithm on a reasonably large (N=1000) validation dataset of dorsal fin images of known individuals, with that of systems primarily designed to aid fin identification in terms of top ranked, and top 10 ranked matches.