@Article{sdhm.2026.078468,
AUTHOR = {Xue Yang, Shunpeng Yang, Wanying Shi, Weizhong Yuan, Sihui Long, Ruixiao Sun, Cheng Yang},
TITLE = {HENet: Hybrid Estimation Architecture with Embedded Physical Constraints for Synergistic Hazy Image Restoration},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/sdhm/online/detail/27367},
ISSN = {1930-2991},
ABSTRACT = {Unmanned aerial vehicle (UAV) imaging techniques have emerged as a promising solution to boost the accuracy and dependability of visual monitoring for railway facilities and peripheral ecological environments, garnering widespread research interest in recent years. Nevertheless, aerial images acquired by UAVs are prone to severe quality deterioration in fog and haze weather scenarios, which greatly hinders the progress and effectiveness of railway routine inspection work. As modern railway systems pursue higher operational safety benchmarks and intelligent rail transit technologies achieve iterative breakthroughs, video monitoring systems have evolved into indispensable core equipment for identifying and early warning of railway traffic abnormal conditions. Targeting the low-quality issue of hazy UAV railway images, this study develops an innovative end-to-end dedicated dehazing neural network. The proposed approach conducts targeted optimization from three critical dimensions: customized network structural design, mixed-type parameter prediction mechanism, and the establishment of a railway-specific hazy image dataset. In particular, a physical prior constraint-based modeling strategy is introduced to analyze the imaging characteristics of foggy and hazy aerial images, which enables the precise restoration of authentic scene structural features and fine texture details of railway infrastructure. Second, a composite loss function is formulated to achieve an improved trade-off among training efficiency, computational cost, and restoration accuracy, thereby accelerating network convergence. Furthermore, image dehazing and physical parameter estimation are seamlessly integrated into a unified framework, wherein mutual collaboration and complementary learning among modules further enhance dehazing performance. Extensive experimental results under the full-reference image quality evaluation paradigm demonstrate that the proposed algorithm outperforms state-of-the-art dehazing methods. It effectively removes haze from railway inspection images, significantly enhances image clarity and contrast, and provides robust technical support and assurance for practical UAV-based railway inspection tasks.},
DOI = {10.32604/sdhm.2026.078468}
}