| Summary: | Haze contamination is a common issue in very high resolution (VHR) remote sensing images, which inevitably hinders their application. The use of auxiliary images has been a key trend for enhancing haze removal. However, auxiliary VHR images are usually inconsistent with the target hazy VHR images, in terms of different imaging viewpoints, misalignment, or land cover changes, which greatly influence the performance of haze removal. To handle this problem, this paper develops a novel solution that incorporates auxiliary VHR images without any preprocessing or manual selection for haze removal. Accordingly, a gradient-based haze removal network (GHRN) is constructed. The gradient maps of both the target hazy images and corresponding auxiliary images are fed to the GHRN architecture, and hierarchical convolutions are applied to assist feature alignment. Seven state-of-the-art haze removal methods were employed as comparisons in ten simulated and six real scenarios. The experiment results demonstrate the reliability of GHRN. Moreover, GHRN is suitable for auxiliary images with different contaminations (e.g., haze or noise). To the best of our knowledge, this is the first study to consider inconsistent auxiliary images for enhancing haze removal.
|
|---|