Two-Stage Generative Color Calibration for Drone Photography With Cloud-Edge Collaboration

There is an increasing demand for the accurate documentation of architectural main colors in the urban color design field. Drone-based photography has emerged as a pivotal tool, due to its capability for high-altitude visual field and convenient data collection. However, a significant challenge remains in ensuring recorded color consistency across varying environmental conditions and timeframes, which requires accurate calibration method. Current calibration methods tend to apply end-to-end neural networks directly on the images, ignoring the difference between the calibration targets of architectural subjects and their surrounding backgrounds. This will lead to significant color deviations in the background areas, resulting in inaccurate calibrated results. Besides, large models are hard to be deployed on computational resource-limited drones. With respect to the above challenges, we propose a novel method called two-stage generative color calibration (TGCC) network for drone photography with cloud-edge collaboration. TGCC tackles the above issues via a two-stage calibration process. The initial stage is conducted at the drone edge side, employing a lightweight neural network for coarse color calibration. Then, the coarsely calibrated images are sent to the cloud server for the subsequent stage, which first extract calibration masks from the target architectural subjects, and then utilizes these masks as guidances for the large generative model to refine color calibration results. Experimental results demonstrate that our approach contributes to higher accuracy and better consistency of color calibration results than prior methods.