Shark identification using the notches in the dorsal fin

Marais, Tessa (2016-12)

Thesis (MSc)--Stellenbosch University, 2016

Thesis

ENGLISH ABSTRACT : In order to protect endangered species, such as the Great White shark, a reliable estimate of the population is needed. In counting animals, it is necessary to be able to distinguish between individuals. The current procedure is manual photo identification, which is a time consuming process. By automating this process, when a shark is spotted, it can easily be matched to an existing shark in the database or added to the database as a new shark. Photo identification software is already available for animal species such as penguins, elephants and dolphins. DARWIN, used for identification of dolphins was tested on sharks, but found to be unsuccessful in matching an unknown shark in to the correct shark in the database. DARWIN also requires extensive user input, which is what we are trying to eliminate or greatly reduce in the identification of sharks. In this thesis, Hidden Markov models (HMM) is used to develop software to identify individual sharks. The results of the HMM was then compared to software using Dynamic Time Warping (DTW) to do the matching. The DTW program was able to correctly match 80% of the images within a rank of twenty and 62% with a rank of two or less. Using HMM, 84% of the photographs were correctly matched with a rank of twenty or less, but only 56% with a rank of two or less and 64% with a rank of five or less. Although the HMM does not perform as well as the DTW, much better performance is expected from the HMM software by building up a quality database through the visual inspection and inclusion of photographs that will lead to more consistent models.

AFRIKAANSE OPSOMMING : In die bewaring van bedreigde spesies, is dit belangrik om 'n betroubare skatting van die populasie te hê. Wanneer die diere getel word, is dit nodig om hulle uit mekaar te kan ken. Huidiglik word foto identifikasie met die hand gedoen, wat 'n baie langdurige proses is. As hierdie proses programmaties gedoen kan word, sou dit moontlik wees om onmiddelik 'n haai uit te ken vanaf 'n foto. Foto identifikasie sagteware is reeds beskikbaar vir pikkewyne, olifante en dolfyne. DARWIN, wat gebruik word vir die identifikasie van dolfyne, was getoets op haaie, maar was onsuksesvol om 'n onbekende haai te identi seer deur dit te pas met die regte haai in die databasis. DARWIN vereis ook ekstensiewe invoer van die gebruiker en dit is juis wat ons probeer uitskakel of verminder in die identifikasie van haaie. In hierdie tesis, word die verskuilde Markov-modelle (HMM) gebruik om sagteware te ontwikkel wat 'n individuele haai kan identifiseer. Die resultate van hierdie sagteware word dan vergelyk met die resultate wat verkry is van die sagteware wat dinamiese tydsverbuiging (DTW) gebruik in die identifisering. DTW kon 80% van die haaivinne korrek identifiseer met 'n rang van 20 of minder en 62% met 'n rang van twee of minder. Hierteenoor kon HMM 84% korrek identifiseer met 'n rang van 20 of minder, maar slegs 56% met 'n rang van twee of een en 64% met 'n rang van vyf of minder. Alhoewel HMM nie so goed soos DTW nie, sal die HMM sagteware heelwat beter vaar wanneer 'n goeie databasis opgebou word deur slegs die mees onlangse, hoë-kwaliteit foto's in die opleiding in te sluit.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/100329
This item appears in the following collections: