Scholarlink

Abstract <p>A system for recognizing and counting two classic Arctic species, the wild reindeer (<i>Rangifer tarandus tarandus</i> L., 1758) and the brent goose (<i>Branta bernicla</i> L., 1758), in photographs from aircraft is presented. The AutoGenNet recognition system is based on a convolutional neural network (CNN) of the Mask R-CNN architecture using the AutoML (Automated Machine Learning) concept. The system uses transfer learning, the essence of which is that at the first stage the system is trained for recognizing a variety of objects applying a standard array of images (about 328 000 images), to be further trained on images of the target objects. This approach allows for a number of images of target objects from several hundreds of thousands in one-stage training to be reduced down to several hundred in two-stage training. A synthesis of the CNN model on the basis of marked images in the AutoGenNet system is performed automatically. A special Markup program was developed for marking animals in photographs and preparing a training sample. The first stage of system training is performed once by SNA and deep learning specialists. The second stage of training can be managed repeatedly in order to retrain the system that made errors in recognizing the objects. The work at this stage can be performed by system users who have no special education in the field of SNS training. Two variants of work with the system are possible: a stand-alone mode in the presence of the necessary computing resources or work via the Internet with the AutoGetNet located on the servers of the St.&#xa0;Petersburg Federal Research Center, Russian Academy of Sciences. The CNN model presented in this contribution was trained based on 100 images of wild reindeer herds. The error of reindeer recognition using an independent data set was about 18%. 260 images of flocks of the brent goose in different environments, be this on land, on water, or in the air, were utilized to recognize Brent geese. The recognition error was about 35%. The AutoGenNet system is unified in terms of recognition objects and can be trained to recognize other animal species without any change in the program, provided they are distinguishable in the images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Monitoring Larger Vertebrates of Arctic Fauna Using Intelligent AutoML Technology

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • V. A. Sobolevskii,
  • L. A. Kolpaschikov,
  • S. B. Rozenfeld,
  • V. V. Mikhailov

摘要

Abstract

A system for recognizing and counting two classic Arctic species, the wild reindeer (Rangifer tarandus tarandus L., 1758) and the brent goose (Branta bernicla L., 1758), in photographs from aircraft is presented. The AutoGenNet recognition system is based on a convolutional neural network (CNN) of the Mask R-CNN architecture using the AutoML (Automated Machine Learning) concept. The system uses transfer learning, the essence of which is that at the first stage the system is trained for recognizing a variety of objects applying a standard array of images (about 328 000 images), to be further trained on images of the target objects. This approach allows for a number of images of target objects from several hundreds of thousands in one-stage training to be reduced down to several hundred in two-stage training. A synthesis of the CNN model on the basis of marked images in the AutoGenNet system is performed automatically. A special Markup program was developed for marking animals in photographs and preparing a training sample. The first stage of system training is performed once by SNA and deep learning specialists. The second stage of training can be managed repeatedly in order to retrain the system that made errors in recognizing the objects. The work at this stage can be performed by system users who have no special education in the field of SNS training. Two variants of work with the system are possible: a stand-alone mode in the presence of the necessary computing resources or work via the Internet with the AutoGetNet located on the servers of the St. Petersburg Federal Research Center, Russian Academy of Sciences. The CNN model presented in this contribution was trained based on 100 images of wild reindeer herds. The error of reindeer recognition using an independent data set was about 18%. 260 images of flocks of the brent goose in different environments, be this on land, on water, or in the air, were utilized to recognize Brent geese. The recognition error was about 35%. The AutoGenNet system is unified in terms of recognition objects and can be trained to recognize other animal species without any change in the program, provided they are distinguishable in the images.