본문 바로가기
  • Home

Characteristics of Loss Functions for Road Detection on Satellite Images

  • Journal of Software Assessment and Valuation
  • Abbr : JSAV
  • 2025, 21(2), pp.51~60
  • Publisher : Korea Software Assessment and Valuation Society
  • Research Area : Engineering > Computer Science
  • Received : May 18, 2025
  • Accepted : June 20, 2025
  • Published : June 30, 2025

Nagyeong Kim 1 Yun Young-Sun 1

1한남대학교

Accredited

ABSTRACT

This study quantitatively analyzes the relationship between loss functions and performance variations in road detection using satellite imagery. Road datasets from AI Hub were utilized, and experiments were conducted based on the NL-LinkNet model, which models long-range feature dependencies. This paper compares and analyzes representative pixel-wise loss functions, Cross Entropy and Focal Loss; region-based loss functions, Dice and Lovasz Loss; and hybrid loss functions combining these. Results showed Dice and Dice-based hybrid loss functions generally provided stable, high-performing results, while Lovasz Loss, though theoretically effective due to its IoU-based formulation, demonstrates poor initial training stability and lower performance in practice. The results suggest that employing hybrid loss functions can provide stable performance and address class imbalance for complex object detection tasks with significant class imbalance, such as road detection. In particular, Dice and Focal-based hybrid loss functions, or Lovasz-based hybrid loss functions, are deemed suitable for balanced performance and initial training stability.

Citation status

* References for papers published after 2023 are currently being built.