|
|
|
 |
Project 0809/7Project Title: Final development of a new computerised fluke matching system and creation of a fluke database for humpback whales photographed off the east coast of Australia from 1999-2005 Chief Investigators: Dr Hendrick Kniest, A/Prof Peter Harrison and Mr Daniel Burns Ongoing development of the computer-aided photo-identification matching system, Fluke Matcher, has resulted in a fully functional version of the program that is able to efficiently identify individuals and find resights in photo-identification catalogues of humpback whales. Fluke Matcher v3.2.2 is superior to previous versions and is now ready to be distributed and used by researchers. The system uses a wide range of criteria, based on multiple key features of humpback whale flukes that are normally utilized by manual matching methods, and additional computerised image-matching techniques, to produce an efficient and reliable matching system. Initial user input is needed to identify control points used for the transformation of the fluke image onto a common reference frame. The system then measures key features of the fluke, including parameters to describe the shape of the fluke, and black and white pigment distribution in different regions of the fluke. The operator then highlights other distinctive features of the fluke that enable identification, such as spots, lines and areas of damage. When searching for a given fluke, the program ranks all of the images in the database from the most likely to the least likely match for that fluke, and the operator then visually checks the ranked images to find the match(es). A database of 440 flukes was entered into Fluke Matcher (FM) and used to develop and refine the program, including improving the data entry phase, identifying which features result in superior and consistent matching results, and identifying the best matching algorithms. Better methods were also introduced to measure the characteristics of the trailing edge and the v-notch area of the fluke, including more information on the shape of these areas. Tests using a number of different operators (both experienced and inexperienced) were conducted to improve the consistency of the input phase of the program, to refine the data-entry protocols, and to improve search methods. Alternative matching algorithms were also developed to make use of existing information within datasets, so that matches that had already been found in previous searches were utilised to remove duplicates from subsequent searches. A Fluke Matcher user manual has also been developed and refined throughout the project. Several sets of pre-matched fluke images were used to test Fluke Matcher. The primary tests were conducted using a set of 860 fluke photographs of 367 whales, with each whale represented by at least two and up to a maximum of five images. An additional 386 images for which there were no matches in the data set were also added to create a very large database of 1247 images (754 whales and 1314 match pairs) with which to test the program. Results of diagnostic tests showed that by using Fluke Matcher (v3.2.2), 48.2% of all matched pairs were found within the top 5 ranked images (i.e. 5/1246 = 0.4%) of the database, and 90.3% of all matched pairs were within the top 120 ranked images (i.e. 120/1246 = 9.6%). A simulated search utilizing advanced search procedures resulted in 98.5% of all matched pairs being found in the top 10% of the ranked database, and 99.5% in the top 20%. These results show that the best results for minimal effort are achieved within the top 10% of the ranked search results. A third data set consisting of 40 images (20 pre-matched pairs) was used to test efficiency when multiple users (two experienced, two neophytes) entered images into Fluke Matcher. This test showed that experienced users were able to consistently enter flukes and efficiently find the matching images entered by other experienced users. Variation in consistency when comparing flukes entered by neophytes suggest that some training is necessary for before new users of Fluke Matcher are able to achieve optimal results, although one neophyte showed strong agreement with the experienced users almost immediately.
|
 
|