PROJECT TITLE :

Quality Prediction on Deep Generative Images

ABSTRACT:

A wide range of applications, including picture production, have recently leveraged Deep Neural Networks. For applications like image reduction, Generative Adversarial Networks (GANs) produce particularly realistic images. Automated evaluation of generative images' perceived quality is desirable for monitoring and controlling the encode process, just like with traditional compression. There are a number of existing image quality algorithms that don't work well with GAN-generated content because of the texture and compression issues. A novel "naturalness"-based image quality predictor for generative images is presented here. Using a multi-stage parallel boosting method, we have developed a novel GAN picture quality predictor based on structural and statistical similarities. In order to facilitate model building and testing, we also created and gathered human views on a subjective GAN picture quality database consisting of (distorted) GAN images. Our experimental results show that our proposed GAN IQA model outperforms traditional image quality datasets when it comes to predicting the quality of generative image datasets.