[HIRING!] 2025 Graduate Students & Undergraduate Interns (석박사 대학원생 및 학부인턴 모집)
Yongrok Kim, Junha Shin, Juhyun Lee, Hyunsuk Ko
To display low-quality broadcast content on high-resolution screens in full-screen format, the application of Super-Resolution (SR) is essential. Recently, SR methods have been developed that not only increase resolution while preserving the original image information but also enhance the perceived quality. However, evaluating the quality of SR images generated from low-quality sources, such as SR-enhanced broadcast content, is challenging due to the need to consider both distortions and improvements. Additionally, assessing SR image quality without original high-quality sources presents another significant challenge. Unfortunately, there has been a dearth of research specifically addressing the Image Quality Assessment (IQA) of SR images under these conditions. In this work, we introduce a new IQA dataset for SR broadcast images in both 2K and 4K resolutions, named Super-Resolution Enhanced Broadcasting contents (SREB) dataset. We conducted a subjective quality evaluation to obtain the Mean Opinion Score (MOS) for these SR images and performed a comprehensive human study to identify the key factors influencing the perceived quality. The SREB dataset and insights from this study can be utilized to develop enhanced IQA metrics that account for scaling factor impacts as well as distortions and improvements commonly encountered in real-world SR applications.
● Y. Kim, J. Shin, J. Lee and H. Ko, "Study of Subjective and Objective Quality in Super-Resolution Enhanced Broadcast Images on a Novel SR-IQA Dataset", arXiv, 2024. [PDF]
● Dataset
Download link: [Google Drive]
● Subjective Evaluation GUI
Download link: [Google Drive]
The SREB dataset includes 30 original low-quality, low-resolution broadcast images, 420 super-resolution (SR) images, and subjective quality scores for SR images. The summary of SREB dataset is given in Table 1. The original images were categorized into three groups: person, background, and screen content, with 10 images in each category, and they exhibit various spatial and color complexities. Seven SR methods were employed, including Bicubic, ASDS, SRCNN, SRGAN, BSRGAN, SwinIR, and SwinIR without deblurring (SwinIR w/o DB). The images were processed using scaling factors of ×2 and ×4, resulting in SR images with resolutions of 1440×1080 and 2880×2160, respectively. A total of 420 SR images (30 images × 7 methods × 2 scaling factors) were generated. Table II summarizes the SR method type, specific method, scaling factor, and the number of images. The Pairwise Comparison (PC) method was employed for the subjective quality assessment of SR images to detect subtle quality differences. The assessment was structured into four sessions: the first two sessions focused on comparisons within the same scaling factor (×2 or ×4), while the final two sessions compared images across the two scaling factors. The results from the first and second sessions were used to compute the Mean Opinion Scores (MOS) for the ×2 and ×4 scaling factors, respectively. The results from all four sessions were combined to compute the MOS for the integrated scaling factors (i.e., ×2×4).
Copyright (c) 2024 Intelligent Visual Media Laboratory (IVML)
All rights reserved.
The SREB dataset is copyrighted by the Intelligent Visual Media Laboratory (IVML). All rights are reserved. Unauthorized use, reproduction, or distribution of this dataset, or any portion thereof, is strictly prohibited without prior written permission from IVML. For any inquiries regarding the dataset, please contact us at helloyr12@hanyang.ac.kr. Permission is not granted, without written agreement and without license or royalty fees, to use, copy, modify, and distribute this dataset (the images and camera calibration data) and its documentation for any purpose.