(C) 2007
Photography of the tail flukes of humpback whales (Megaptera novaeangliae) is the primary method of identifying individual animals, enabling the illumination of life history, migration, and social behavior. Matching of photographs is currently performed manually, generally by comparing pigmentation patterns on the ventral face of the fluke. However approximately one third of whales possess uniformly pigmented (all-black or all-white) flukes, which necessitates matching be performed on the scalloped pattern of the fluke’s trailing edge contour, a particularly labor-intensive process prone to fatigue and error. The development of a semi-automated computer-based system for the matching flukes via edge contours is presented.
In the pilot experiment, location features were extracted from the edge contour with a discrete wavelet transform (DWT) with Symlet 5 as the basis wavelet; in subsequent experiments, a continuous wavelet transform (CWT) using the same basis wavelet was utilized. Feature matching was performed with a knowledge-based system emphasizing the matching of clusters of features within local subsets of the full contour, reflecting the manual matching strategies of experts. A set of 250 photographs of all-black flukes representing an entire field season’s collection from the research program of The Dolphin Institute (TDI) was used for testing and development of the system.
A hit rate of 62% and a false alarm rate of 0.16% were achieved, so that approximately one in every three match suggestions was a true match. This hit rate is comparable to estimates of the performance of a single pass of traditional manual matching on these all-black flukes; in TDI’s protocols, three such passes are ordinarily performed. These results validate the utilization of location features as well as knowledge-based matching based on localized clusters of these features, and may justify the replacement of a single pass of manual matching with the computer-assisted system. Future development will be directed toward usability requirements such as semi-automated edge extraction and an end-user interface, as well as efforts to further improve the hit rate so that the computer-assisted system may fully replace manual matching.
Dissertation committee:
Louis M. Herman, Chairperson
Todd R. Reed, Outside member (Dep't of Electrical Engineering)
Karl A. Minke
Joseph R. Mobley, Jr.
Adam A. Pack
