No-Reference Quality Assessment for Pansharpened Images via Opinion-Unaware Learning

The high-quality pansharpened image with both high spatial resolution and high spectral fidelity is highly desirable in various applications. However, existing pansharpening methods may lead to spatial distortion and spectral distortion. To measure the degrees of distortion caused by the pansharpening methods, we conduct in-deep studies on the subjective and objective quality assessment of pansharpened images. We built a subjective database consisting of 360 images generated from 20 couples of panchromatic (PAN)/multispectral (MS) images using 18 pansharpening methods. Based on the database, we proposed a no-reference quality assessment method to blindly predict the quality of pansharpened images via opinion-unaware learning. The proposed method first extracted features from the MS images' spectral bands and typical information indexes which comprehensively reflect spatial distortion, spectral distortion, and the effects of pansharpening on applications. Based on the features extracted from the pristine MS image training dataset, a benchmark multivariate Gaussian (MVG) model is learned. The distance between the benchmark MVG and the MVG fitted on the test image is calculated to measure the quality. The experimental results show the superiority of our method on our database.