YOLO-Based Object Localization and Classification in UAV Images Compressed by JPEG

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Methods for object localization and classification in images acquired from unmanned aerial vehicles (UAVs) quickly develop and find new applications. Pre-trained convolutional neural networks (CNNs) play the key role in solving these tasks. However, there are many factors that degrade the quality of acquired images and make the performance of methods intended for object detection and classification worse. One such factor is lossy compression of acquired images or video data widely used to pass them from on-board sensors and devices of preliminary data processing to on-land centers that perform further data processing for retrieval of valuable information. Both CNNs applied for localization and classification, and lossy compression techniques used to reduce the transferred data size have an impact on final results. To study this impact, we analyze the performance of several modifications of YOLO (You Only Look Once) CNNs applied to color images compressed by JPEG, which continues to be one of the basic compression tools. The quality factor is varied within wide limits to detect the situation when distortions due to lossy compression start to become too large and have a considerable negative effect on the localization and classification of objects of different types and sizes. Analysis is carried out using several traditional criteria, including Intersection over Union, F1, and mAP metrics, as well as some others. Dependence of localization and classification characteristics on the object size is performed. The datasets VisDrone and TAI are employed in training and verification.