LCEA-YOLO: Improving Small Object Detection in Aerial Imagery Using Local Contrast Enhancement Attention and Inner-CIoU Loss

Résumé

This study proposed the Local Contrast Enhancement Attention - You Only Look Once (LCEA-YOLO) model to address the challenge of detecting small objects in high-altitude Unmanned Aerial Vehicle (UAV) images. LCEA-YOLO is a customized detector developed using the YOLOv10n architecture that embeds an LCEA module to emphasize local subtle contrast variations, as well as an Inner-CIoU loss function that enhances the regression ability for smaller objects. Experiments on the UAVDT dataset were conducted, with the results demonstrating that LCEA-YOLO exceeded other baseline algorithms with an overall mAP@0.5 of 47.2% and mAP@0.5:0.95 of 29.9%. Additionally, the model presented remarkable resilience on challenging under-represented categories, yielding improvements of 15.0% and 7.6% in detection accuracy for trucks and buses, respectively. These findings validated the benefit of targeted local contrast enhancement and scale-aware regression for real-time UAV detection.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0