Litcius/Paper detail

Enhanced YOLOv8 for Small Object Detection in UAV Aerial Photography: YOLO-UAV

Hongcheng Xue, Xinyi Wang, Yuantian Xia, Zhan Tang, Lin Li, Longhe Wang

202412 citationsDOI

Abstract

Detecting objects in images captured by Unmanned Aerial Vehicles (UAVs) is pivotal across civilian, commercial, and military domains. Nevertheless, prevalent embedded detection models for UAVs face accuracy challenges, particularly with the prevalence of small-sized objects and constrained hardware resources. To tackle these issues, we have enhanced YOLOv8, introducing a specialized object detection model tailored for UAV aerial scenes, named YOLO-UAV. Firstly, we introduce a specialized layer meticulously crafted for the enhanced detection of small objects. This layer serves to elevate the fusion pathway between deep and shallow features, mitigating the incidence of false negatives in small object detection and consequently elevating the overall detection capabilities. Secondly, we introduce a feature processing module, designated as C2f-EMBC, engineered to facilitate an exhaustive fusion of both shallow and deep features. Lastly, to address the model’s vulnerability to positional deviations of small-sized objects, we employ the normalized Wasserstein distance (NWD) as a accurate measure for enhancing small object detection precision. The outcomes demonstrate a substantial enhancement in forecasting precision compared to the baseline model.

Topics & Concepts

Aerial photographyObject detectionComputer scienceComputer visionObject (grammar)PhotographyArtificial intelligenceRemote sensingComputer graphics (images)GeographyArtVisual artsSegmentationAdvanced Neural Network ApplicationsInfrared Target Detection MethodologiesRobotics and Sensor-Based Localization