Blind Quality Assessment Method to Evaluate Cloud Removal Performance of Aerial Image

People often use image-inpainting-based methods to remove cloud from aerial images, but it lacks a targeted quantitative evaluator to assess the removal result. In order to solve this issue to some extent, we propose an assessment method that combines ranking learning and regression learning. The main framework consists of several CNN down-sampling steps layer-by-layer. Firstly, to learn the distortion features of the inpainted image, an image classification task is conducted to train the network. Secondly, we use the proposed joint loss function to regress the features into the FR-IQA evaluator SSIM by retraining the whole network. Through end-to-end training, the proposed model learns a priori of the aerial image and realizes the approximation of SSIM. Experimental results demonstrate that our method has achieved better performance on SROCC, RMSE and PLCC in comparison with other blind image quality assessment methods.