DEHAZING AERIAL DRONE IMAGES VIA REGIONAL SATURATION-VALUE MAPPING AND U-NET-DRIVEN SOFT SEGMENTATION AND YOLO DETECTION
DOI:
https://doi.org/10.62643/Abstract
Aerial drone imagery plays a critical role in diverse applications such as surveillance, environmental monitoring, and mapping. However, haze and atmospheric distortions often degrade image quality, reducing contrast, color fidelity, and visibility of critical details. To address this challenge, we propose a hybrid dehazing framework that integrates Regional Saturation-Value (SV) Mapping with a UNet-driven Soft Segmentation module, further enhanced by a pretrained YOLObased object detection network. The first stage of our method analyzes regional variations in saturation and value within the HSV color space to adaptively enhance hazeaffected regions, improving local contrast and compensating for spatially varying haze density. Next, a pretrained YOLO model is employed to detect key objects and regions of interest (ROIs) within the aerial scene, guiding the subsequent segmentation process to focus selectively on object-rich and detailcritical areas. In the second stage, a modified U-Net architecture performs soft semantic segmentation within these YOLO-guided ROIs, refining object boundaries and background textures while preventing overenhancement or artifacts. This synergistic combination of localized color-space enhancement, object-aware detection, and deep segmentation ensures both structural preservation and perceptual realism. Experimental evaluations on benchmark aerial image datasets demonstrate that our framework surpasses existing dehazing techniques in PSNR, SSIM, and perceptual quality. Moreover, the integration of pretrained YOLO detection enhances the interpretability of restored scenes and improves the performance of downstream computer vision tasks such as object tracking,target recognition, and scene understanding
in aerial imagery.
Downloads
Published
Issue
Section
License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.













