A CNN-Based Quality Assessment Method for Pseudo 4K Contents

Abstract

AbstractRecently, there has been a growing interest in Ultra High-Definition (UHD) content, which brings a better visual experience for end-users. However, quite a few contents with 4K resolution are upscaled from High-Definition (HD) contents and suffer degradations in quality, such as blur, texture shift, etc. These pseudo 4K contents can not deliver the expected quality of experience (QoE) to end-users while requiring a high transmission bit rate in the meantime, which inevitably results in a waste of bandwidth resources. Hence, we develop a novel deep learning-based no reference (NR) image quality assessment (IQA) model for recognition and quality evaluation of real and fake 4K images. To reduce the computational overhead for a 4K image, we first select three representative patches with high texture complexity by the Grey-Level Co-occurrence Matrix (GLCM) based measure. Next, the convolutional neural network (CNN) is adopted to extract the quality-aware features of three representative patches. Specifically, we extract different levels of features from intermediate layers of CNN and concatenate them into a more effective quality-aware feature representation. Finally, the shallow fully connected (FC) network is utilized to aggregate the features into the quality score and the overall quality score of the 4K image is calculated as the average value of three patches’ quality scores. The experimental results show that the proposed method outperforms all compared NR IQA metrics on the 4K IQA database.KeywordsImage Quality Assessment (IQA)Pseudo 4K contentNo Reference (NR)Deep learning