A Neural Network Based Automated Identification System For Biological Species

该奖项为开发一种完全自动化的系统提供支持，该系统能够根据标本的数字图像提供物种级别的生物识别。该技术使用人工神经网络，可以使用通过互联网传输的图像作为输入。目前，获得物种一级的标本鉴定可能是一项困难的工作。对大多数生物体群体进行系统化训练的人数很少；试图实现准确鉴定的非专家人员的成功率往往较低。这样的系统有可能从根本上增加生物多样性信息在保护和科学方面的使用。该项目旨在开发能够以蜘蛛为测试对象进行物种识别的软件技术。这种方法是基于通过配备相机的解剖显微镜拍摄的生物体分类相关结构的数字图像。对图像进行小波变换，这是一种从图像中提取形状信息，同时去除不太有用的高分辨率信息的过程。通过这个过程，图像被简化为一组小波系数，可以提供给称为人工神经网络的计算算法。这样的网络能够学习对物体进行分类。该项目的目标是建立一个识别准确率为95%的系统。在该项目的第一阶段，将使用三个数据集：澳大利亚蜘蛛的两个家族(斑点蛛科和原蜘蛛科)的图像，以及第三组图像，这些图像是从地理而不是分类的角度定义的。这一地理集合包括连续三年在田纳西州选定地点收集的标本，是典型的生态调查数据。在第二阶段，将开发一个网络界面，允许通过互联网提交图像进行识别。目前，严重限制我们理解群落结构、多样性以及多样性如何与生态系统功能和由此产生的人类服务相关的一个因素是缺乏能够向物种识别生物标本的专家。因此，自动识别的潜在影响是巨大的。一个系统可以识别特定科或特定地区的任何物种，而不需要使用者对要识别的生物体有更多的最基本的知识，这样的系统有可能极大地提高生物清单和后续监测工作的效率和范围。

Abstract0119578PlatnickThis award provides support for the development of a completely automated system capable of providing species-level identifications of organisms based on digital images of specimens. The technology uses an artificial neural network and can use, as input, images transmitted over the internet. At present, obtaining species-level identifications of specimens can be a difficult undertaking. The number of trained systematizes for most groups of organisms is low; the success rate of non-specialists trying to achieve accurate identifications is often lower. Such a system has the potential to radically increase the use of biodiversity information in conservation as well as science. This project is aimed at developing software technology capable of species identification using spiders as test subjects. The approach is based on digital images of taxonomically relevant structures of the organism taken through a camera-equipped dissecting microscope. The images are subjected to wavelet transformation, a procedure that extracts shape information from the image while removing less useful, high-resolution information. By this procedure, the image is reduced to a set of wavelet coefficients that can be supplied to a computing algorithm known as an artificial neural network. Such networks are capable of learning to classify objects. The goal is creation of a system with a 95% accuracy in identification.In the first phase of the project, three datasets will be used: images of two families of Australasian spiders (Trochanteriidae and Prodidomidae), and third set of images defined geographically rather than taxonomically. This geographic set includes specimens from three consecutive years of collecting at selected sites in Tennessee, and is typical of ecological inventory data. In the second phase, a web interface that allows submission of images for identification over the internet will be developed.At present, a severely limiting factor on our understanding of community structure, diversity, and how diversity relates to ecosystem function and resulting human services is the lack of experts capable of identifying biological specimens to species. Thus, the potential impact of automated identification is enormous. A system that can identify any species in a particular family, or from a particular area, without requiring the user to have more than the most basic knowledge of the organism to be identified, has the potential to drastically improve the efficiency and scope of biological inventories, and subsequent monitoring efforts.