Detection and classification of small-size flying objects in images using convolutional neural networks from the YOLOv5 family

Authors: Klekovkin V. A., Markov N. G., Nebaba S. G.

Annotation: The effectiveness of two models of convolutional neural net-works, YOLOv5s and YOLOv5x of the YOLOv5 family, selected as a result of the analysis of the main known models of such networks for solving the problem of detection and classifi-cation of small-sized flying objects in images, was studied. Two datasets were formed with labeled RGB images of flying ob-jects for three classes: birds, rotary-wing unmanned aerial vehi-cles (UAVs), and fixed-wing aircraft. The first dataset contains images with small-size objects (up to 32x32 pixels in area), and the second dataset, used for comparative studies, contains imag-es with objects of different size. The network models were trained using each of the datasets and then studied using their test samples. It was found that both network models, regardless of the object size category, show a computation speed signifi-cantly exceeding the threshold value of the FPS metric equal to 25. It was shown that in the case of small-size flying objects, the YOLOv5s model does not reach the required threshold value of 0.9 in terms of classification accuracy according to the AP0.5 metric for objects of the «Bird» class, while the YOLOv5x model exceeds it and can be used in computer vision systems. For flying objects of different size, both models meet the formulated requirements for the accuracy of object classification in images and for the speed of their computation and are recommended for use in real-time computer vision systems.