Deep Feature Map-Driven No-Reference Quality Assessment For Satellite Images

The practical value of remote sensing images depends heavily on their quality, but they often suffer from distortions such as noise and blur from the imaging process onwards. Some full reference image quality assessment (FR-IQA) methods attempt to create metrics that align well with the human visual system (HVS), but they require high-quality reference images, the lack of which in real-world scenarios makes IQA a challenging task. Hence, there is a need for an adequate no-reference (NR) metrics that do not require reference images. This work focuses on predicting image quality in a NR setting using a learning-based NR-IQA approach to estimate popular measures like SSIM without a reference. We design an end-to-end deep IQA network to compute distortion maps and aggregate a quality score for an image. We create a custom benchmark dataset of high-resolution satellite images with varying distortions to demonstrate excellent performance of our algorithm in our experiments.