IDBR: Development of Tools for Individual Recognition of Animals

达特茅斯学院的道格拉斯·博尔格博士和汉尼·法里德博士获得了一笔赠款，用于开发个人识别动物的工具。在空间和时间上识别和跟踪单个动物的能力可能是动物种群生物学最重要的工具。通过识别个体，研究人员可以估计人口规模、出生率和死亡率，并量化社会行为。这些参数构成了最纯粹和最实用的种群生物学的基础。传统上，这种识别是通过捕捉动物并在它们身上留下可见和独特的标记来实现的。这些方法被称为标记-再捕获或标记-再对准。使用传统标记技术的主要限制是动物福利、成本和难度。一种很有前途的非侵入性技术是使用照相标记？和回顾方法。对于有独特标记的动物，可以给个体拍照(标记)，并将图像存储在数据库中。然后，可以将后来拍摄的动物与图像数据库进行比较，以确定该个体是否以前被看到过(重新审视)，或者它是否是该研究的新成员。这种方法被手动用于研究相对较小的种群，如鲸鱼。为了在大种群中使用，这种图像匹配过程需要计算机辅助才能可行。我们的项目是生物学和计算机科学的合作，开发和测试用于动物个体识别的开源应用程序。该应用程序将包括以下模块(1)用于存储和访问单个图像的图像数据库；(2)几种模式提取技术的选择；以及(3)几种模式匹配算法的选择。该系统将处理单个动物的数字照片，高效地提取必要的图案信息，将这些信息存储在数据库中，并高效地在现有数据库中搜索匹配的图像。要实现的特征检测器包括Harris检测器、可控滤波器和尺度不变特征变换(SIFT)。模式识别算法将包括最近邻、主成分分析、线性判别分析和K-均值。该系统将与我们现有的来自坦桑尼亚北部塔兰盖尔生态系统的8,250张角马图像的mark-resight照片数据库进行测试。此外，我们将拍摄来自同一生态系统的另外两种具有独特图案的有蹄类动物的图像：斑马和长颈鹿。我们将使用这些数据和分析来估计角马的种群规模、生存和招募。对于长颈鹿和斑马，我们将能够估计种群数量。除了这三种非洲有蹄类动物外，我们还将在两个合作伙伴的斑点火蜥蜴和鲸鲨图像数据库中测试这个系统。系统性能的评估将基于它对这些测试数据库产生的误识率，以及系统对不同物种使用的一般适应性。工具将使用常见的开源应用程序开发，并将通过达特茅斯学院-S网站提供给其他研究人员。近几年来，标记重现数据的分析方法有了很大的进步。这些新的分析技术允许更准确的参数估计，并使研究人员能够严格测试复杂的假设。然而，这些方法的应用受到了相对较少的人口数量的限制，这些人口拥有足够的重测数据。我们创建的工具的可用性应该会导致可以使用摄影标记-重视方法进行监测和参数化的人口数量至少增加一到两个数量级。这反过来应导致对这些动物种群进行更知情的管理和保护。

A grant has been awarded to Drs. Douglas Bolger and Hany Farid of Dartmouth College to develop tools for the individual recognition of animals. The ability to recognize and follow individual animals over space and time is perhaps the most important tool of animal population biology. Recognizing individuals allows researchers to estimate population size and birth and death rates, and quantify social behavior. These parameters form the basis of most pure and applied population biology. Traditionally, this recognition has been accomplished by capturing animals and placing visible and unique marks on them. These methods are known as mark-recapture or mark-resight. The primary limitations on the use of traditional marking techniques are animal welfare, cost and difficulty. One promising non-invasive technique is the use of photographic ?mark? and resight methods. For animals with unique markings, individuals can be photographed (marked) and the images stored in a database. Animals photographed later can then be compared to the image database to determine if that individual had been seen before (a resight) or if it is new to the study. This method has been used manually for the study of relatively small populations such as those of whales. For use in large populations this image matching process needs to be computer-assisted to be feasible.Our project is a collaboration between biology and computer science to develop and test an open-source application for individual recognition of animals. This application will include the following modules (1) an image database for the storing and accessing of individual images; (2) several choices of pattern extraction techniques; and (3) several choices of pattern matching algorithm. This system will process digital photographs of individual animals, efficiently extract the essential pattern information, store this information in a database, and efficiently search the existing database for matching images. The feature detectors to be implemented include the Harris detector, steerable filters, and scale invariant feature transform (SIFT). Pattern recognition algorithms will include nearest neighbor, principal components analysis, linear discriminant analysis, and K-means. The system will be tested against our existing mark-resight photographic database of 8,250 wildebeest images from the Tarangire ecosystem in northern Tanzania. Furthermore, we will capture images of two other uniquely patterned ungulates from the same ecosystem: zebra and giraffe. We will use these data and analyses to estimate population size, survival and recruitment for wildebeest. For giraffe and zebra we will be able to estimate population size. In addition to these three African ungulate populations we will also test this system against two collaborator image databases of spotted salamanders and whale sharks. System performance will be evaluated on the basis of the misidentification error rates that it produces for these test databases, as well as the general adaptability of the system for use with different species.The tools will be developed using common open source applications and will be made available to other researchers through Dartmouth College?s website. In recent years there have been tremendous advances in analytical methods for mark-resight data. These new analytical techniques allow for more accurate parameter estimation and give researchers the ability to rigorously test complex hypotheses. However, the application of these methods has been limited by the relatively small number of populations that have sufficient mark-resight data available. The availability of the tools we create should lead to an increase in at least one to two orders of magnitude in the number of populations that can be monitored and parameterized using photographic mark-resight methodology. This should in turn lead to more informed management and conservation of these animal populations.