Technology development for determining tree species using computer vision

Abstract

A technology has been developed to determine the European white birch (Betula pendula Roth.) species in the photo. The differences of the known neural networks of classifiers with the definition of objects are studied. YOLOv4 was chosen as the most promising for further development of the technology. The mechanism of image markup for the formation of training examples has been studied. The method of marking on the image has been formed. Two different datasets have been formed to retrain the network. An algorithmic increase in the dataset was carried out by transforming images and applying filters. The difference in the results of the classifier is determined. The accuracy when training exclusively on images containing hanging birch was 35 %, the accuracy when training on a dataset containing other trees was 71 %, the accuracy when training on the entire dataset was 75 %. To demonstrate the work, birch trees were identified in photographs taken in the arboretum of the MF Bauman Moscow State Technical University. To improve the technology, additional training is recommended to determine the remaining tree species. The technology can be used for the implementation of taxation of specific tree species; the formation of marked datasets for further development; the primary element in the tree image analysis system, to exclude third-party objects in the original image.