Predicting the quality of images compressed after distortion in two steps

摘要: Full-reference and reduced-reference image quality assessment (IQA) models assume a high reference against which to measure perceptual quality. However, this assumption may be violated when the source is upscaled, poorly exposed, or otherwise distorted before being compressed. Reference IQA on compressed but previously “reference” produce unpredictable results. Hence we propose 2stepQA, integrates no-reference (NR) (R) measurements into prediction process. The NR module accounts for imperfect of image, while R component measures further from compression. A simple, efficient multiplication step fuses these single score. We deploy MS-SSIM as NIQE combine them using multiplication. chose MS-SSIM, since it correlates well with subjective scores. Likewise, efficient, generic, does not require training data. 2stepQA approach can generalized by combining other models. also built new data resource: LIVE Wild Compressed Picture Database, where authentically images were JPEG at four levels. shown achieve standout performance compared proposed made publicly available https://github.com/xiangxuyu/2stepQA.