Quality Prediction on Deep Generative Images

Increasing Video Perceptual Quality with GANs and Semantic Coding

“…As expected, and noted in other works that applied generative approaches to image quality improvement like [13,23], the SSIM metric shows a small decrease in both cases. This is due to the fact that GANs "hallucinate" details, thus signal-based full-reference metrics are unable to account for the improvements.…”

“…This is in line with the copious amount of results that shows how images ranked higher by humans obtains a lower score according to SSIM and PSNR metrics. For these reasons a lot of work has been dedicated to obtain more reliable metrics for image quality assessment [23,32,33,48]. In our work we will rely on modern LPIPS metric as a full-reference evaluation and BRISQUE for a no-reference image quality assessment.…”

“…The SSIM index varies between -1 and 1, where 1 indicates perfect structural similarity, while 0 indicates no structural similarity. However, it must be noted that, as reported in [23], many existing image quality algorithms like SSIM are unreliable on GAN generated content, since images generated by GANs may appear quite realistic and similar to an original, yet may match it poorly based on simple pixel comparisons; metrics based on "naturalness" are more suitable in this case. Nevertheless, we report results using this metric due to its widespread use.…”

Fast Video Visual Quality and Resolution Improvement using SR-UNet

“…We notice that our architecture largely improves the perceptual metric (LPIPS) over H.265, thanks to the compression artifacts reduction and to the increase of frequencies in the high-frequency spectrum, and the quality improvement is also notable by the increase in VMAF. The SSIM metric, although being more perception-oriented than PSNR, is still based on the original signal, thus it is somehow predictable how its score is reduced by the adversarial and perceptual-driven training, as reported also in [18]. SR-UNet obtains a large speed-up over the SR-ResNet while maintaining the same quality.…”

“…These metrics are particularly relevant when evaluating generative models since. In fact, as reported in [18] following a subjective study, images generated by GANs may appear quite realistic and similar to an original, yet they may match it poorly based on simple pixel comparisons; metrics based on "naturalness" are thus more suitable in this case.…”