Comparative Analysis of the Vehicle Recognition Method’s Effectiveness on Aerial Images

Introduction. Object recognition on aerial images is an urgent task in modern conditions, especially in cases requiring accurate and fast car recognition. Traditional contour extraction methods, such as Canny, Sobel, Laplacian, Prewitt, and Scharr, are based on gradient analysis and are known for their ease of implementation. This is an essential step for further recognition, as the correct definition of contours contributes to more accurate object identification. However, the effectiveness of the above methods could be improved, especially in complex environments with high object density, uneven brightness, and noise. Neural network models, such as YOLO (You Only Look Once), offer new possibilities, providing more accurate and reliable recognition, even in difficult situations. Purpose. This study compares the effectiveness of classical contour extraction methods and the YOLOv6n neural network model for vehicle recognition in aerial images. The accuracy of vehicle detection is evaluated by the main metrics: Precision, Recall, and F1-measure, which allow the determination of each method’s efficiency level in specific conditions. Methods. The study includes testing the classical Canny, Sobel, Laplacian, Prewitt, and Scharr algorithms for car outline detection and analyzing the results of the YOLOv6n model for deep-learning object detection. Classical methods use image processing to identify contours based on pixel gradients, which allows for extracting structures in an image. The YOLOv6n model is based on a neural network approach, considering complex image features for more accurate and faster object detection. Results. The data analysis showed that classical methods, although they can detect contours, have limited accuracy in conditions of high object density and sharp changes in brightness. The accuracy (Precision) and F1 Score for traditional methods was low, indicating a significant number of false positives and false negatives. In particular, the Sobel and Scharr methods showed the highest Recall but significantly decreased accuracy. In contrast, the YOLOv6n neural network model demonstrated high results in all primary metrics: Precision – 97.9%, Recall – 94.8%, F1 Score – 96.32%, and maP – 97.6%, which confirms its advantages in providing accurate and reliable vehicle recognition in aerial images. Conclusions. The study has shown that traditional contour extraction methods can serve as auxiliary tools for image preprocessing. Still, they need to provide adequate accuracy for the final stages of vehicle recognition. Neural network approaches, such as YOLOv6n, significantly outperform classical methods by providing high detection speed and accuracy, making them recommended for use in high-precision object recognition tasks in aerial images.