The deep colour consistency correction network for remote sensing images

Radiometric discrepancies in remote sensing images, induced by seasonal variations, illumination changes and atmospheric disturbances, often result in prominent stitching artefacts in mosaic images. These artefacts degrade the visual consistency of the images and hinder the accurate interpretation of information. Existing methods rely primarily on traditional radiometric correction models, which are limited in their generalization across diverse image types due to the scarcity of suitable datasets. Therefore, we propose FB-Correction, a dual-branch colour consistency correction model for large-scale remote sensing images. The method performs high-precision correction at low resolutions while enabling low-error parameter mapping at high resolutions. In the high-resolution branch, pixel-level colour transformation is applied to interpolate network outputs in the parameter space, reducing interpolation errors in the image space. The low-resolution branch employs GLNet, designed to capture both short-range and long-range dependencies. To support training, we develop an efficient data collection method based on change detection and semantic segmentation datasets, constructing a task-specific dataset for colour consistency correction in remote sensing images. Experimental results across multiple datasets demonstrate that our method achieves realistic and harmonious image correction. Additionally, it outperforms baseline models across various metrics while attaining the lowest inference time and FLOPs in colour consistency correction tasks, validating its strong generalization capability.