DETECTION OF DEER IN IMAGES TO ACCOUNT POPULATIONS BY COMPUTER VISION METHODS

Abstract

Purpose. Recognition of deer individuals by their graphic images based on convolutional neural network models. Research methods and materials. The material for the research was marked images of deer in different habitats, published in the public domain. Research methods: theory of design and development of artificial intelligence systems; image augmentation for computer vision tasks; hyperparameter tuning algorithms for training neural network models. Results. Conservation of the deer population is one of the important goals in ecology and agriculture. The control of the number of individuals affects the conservation of rare species of deer, and supports the production volumes on farms. For example, such a control method may be the detection of deer individuals in photographs. To accurately determine and count the number of deer, you can use a tool such as a neural network. We used deep learning methods for convolutional neural networks, as well as the concept of “transfer learning”. On the basis of the Faster R-CNN Resnet50 network, a neural network was trained, which allows, with an accuracy of 0.91, to determine individual individuals in the images using the F1-score metric with a threshold value of 0.6 (coincidence of the area of the predicted markup and the actual one) with an accuracy of 0.91. Conclusion. To solve the problem of deer detection, a data set has been prepared, including more than 30 thousand images, with markings of individual individuals. The marking of each individual included the coordinates of the bounding rectangle in the image. Based on this information, a neural network model was developed and trained using software tools to solve the problem of detecting deer objects in images. The experiments carried out showed that the accuracy of detection with augmentation in accordance with the F1-score indicator with a threshold value of 0.6 on the training sample was 0.96, and on the test sample was 0.91.

Keywords

• graphic images
• reindeer husbandry
• augmentation
• neural networks
• artificial intelligence
• object detection