| Title |
Research on Generating a Domain Gap Reduction Hyperspectral Fog Synthesis Dataset through Modeling Wavelength-Specific Scattering Characteristics from a Frontal View |
| Authors |
박정은(JungEun Park) ; 허석행(Seokhaeng Heo) ; 조용진(Hyewon Bang) ; 방혜원(Yongjin Jo) ; 김성호(Sungho Kim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.10.29 |
| Keywords |
Hyperspectral imaging; Haze synthesis; Front-view imaging; Spatial?spectral transmission; Dehazing |
| Abstract |
Hyperspectral imaging (HSI) captures fine-grained spectral information over a broad wavelength range. However, atmospheric scattering caused by haze and aerosols degrades image contrast and signal intensity and distorts spectral information, consequently reducing the performance of recognition algorithms. Training deep learning-based hyperspectral image dehazing models requires a large number of paired hazy?clean images acquired from identical scenes; however, collecting such data in real-world environments is challenging. Moreover, conventional RGB-based haze synthesis methods do not adequately account for wavelength-dependent scattering characteristics, resulting in a domain gap between real and synthetic hazy images. To address these limitations, this study moves beyond conventional statistically estimated parameters and proposes a front-view spatial?spectral haze synthesis framework that integrates measured airlight with wavelength-dependent transmission characteristics. Using 519-band hazy?clean image pairs acquired with a SPECIM line-scan hyperspectral camera, wavelength-dependent airlight is estimated from a black velvet reference target, while the spatial haze distribution is derived using a lower-percentile-based method. These components are combined to construct a three-dimensional spatial?spectral transmission cube, and haze images with varying densities are generated using a density-adjustment coefficient. The normalized spectral profiles of real and synthetic haze achieved a spectral correlation of 0.9347, and the proposed method reproduced pixel distributions closer to those of real hazy images than two ablation variants. Furthermore, dehazing models trained on the proposed synthetic dataset achieved lower average restoration errors on real hazy images, demonstrating a reduced domain gap and improved generalization to real-world conditions. |