EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Detector Redesign and Distance-Corrected CIoU Loss for YOLO11-Based Face Detection

Aylın Şevık Yalçın, Pakize Erdoğmuş

IEEE Access · 2026 · 14 · 103645-103665

DOI: 10.1109/access.2026.3706494 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Aylın Şevık Yalçın und 1 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Face detection remains a challenging problem in unconstrained environments because of scale variations, occlusions, pose changes, dense crowd layouts, and the need for real-time inferences. This study investigated face detection from a joint architectural and localization perspective by redesigning a YOLO11n-based detector and introducing displacement-aware regression loss. Specifically, alternative ShuffleNetV2- and CSPResNeXt-based backbone-neck configurations were integrated into the detector while maintaining the same three-scale detection paradigm, and a Distance-Corrected Complete Intersection over Union (DC-CIoU) loss was proposed to moderate overlap penalization under structured spatial shifts. Using the baseline model as a reference, 13 additional detector variants were evaluated in a controlled setting to analyze the individual and combined effects of architectural redesign and loss reformulation. The proposed framework is assessed on the WIDER FACE for primary evaluation and the Face Detection Data Set and Benchmark (FDDB) for cross-dataset analysis, together with deployment-oriented benchmarking and video-based runtime experiments across multiple inference formats. The results showed that CSPResNeXt-based variants achieved the strongest overall detection accuracy, whereas the Distance Corrected Complete Intersection Over Union consistently improved recall-oriented behavior and hard-subset performance under challenging face detection conditions. Additionally, runtime analysis demonstrated that YOLO11n remains the optimal baseline model in terms of efficiency, whereas TensorRT provides the fastest deployment backend. Overall, the findings indicate that DC-CIoU is a task-oriented localization refinement mechanism whose practical benefit becomes more evident when combined with suitable detector architectures in challenging face detection scenarios.

Zitieren

APA

Yalçın, A. Ş., & Erdoğmuş, P. (2026). Detector Redesign and Distance-Corrected CIoU Loss for YOLO11-Based Face Detection. IEEE Access, 14, 103645-103665. https://doi.org/10.1109/access.2026.3706494

Vancouver

Yalçın AŞ, Erdoğmuş P. Detector Redesign and Distance-Corrected CIoU Loss for YOLO11-Based Face Detection. IEEE Access. 2026;14:103645-103665. doi:10.1109/access.2026.3706494

BibTeX
@article{Yaln2026,
  author = {Yalçın, Aylın Şevık and Erdoğmuş, Pakize},
  title = {Detector Redesign and Distance-Corrected CIoU Loss for YOLO11-Based Face Detection},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {103645--103665},
  doi = {10.1109/access.2026.3706494}
}
Kostenlos starten →