An improved objective metric to predict image quality using deep neural networks

Abstract

Objective quality assessment of compressed images is very useful in many applications. In this paper we present an objective quality metric that is better tuned to evaluate the quality of images distorted by compression artifacts. A deep convolutional neural networks is used to extract features from a reference image and its distorted version. Selected features have both spatial and spectral characteristics providing substantial information on perceived quality. These features are extracted from numerous randomly selected patches from images and overall image quality is computed as a weighted sum of patch scores, where weights are learned during training. The model parameters are initialized based on a previous work and further trained using content from a recent JPEG XL call for proposals. The proposed model is then analyzed on both the above JPEG XL test set and images distorted by compression algorithms in the TID2013 database. Test results indicate that the new model outperforms the initial model, as well as other state-of-the-art objective quality metrics.