Super-resolution (SR) methods can amplify or reduce distortions in low-quality broadcast content, and these SR-induced distortions and enhancements impact the quality of the resulting SR images. Additionally, as the scaling factor increases, SR image quality tends to deteriorate. Evaluating SR image quality in real-world applications is particularly challenging due to the absence of high-quality, high-resolution originals for comparison. Therefore, research on SR image quality assessment (SR-IQA) must carefully consider these factors. In this work, we introduce a novel dataset named SRCD, specifically designed for contrastive learning. This dataset includes distorted images and their corresponding SR images, generated using various SR methods and scaling factors. It enables a deep learning-based encoder to learn SR-induced distortions and enhancements, as well as the effects of scaling factors.

This dataset is provided by the Intelligent Visual Media Laboratory (IVML) and is intended for research and academic use only. By using this dataset, you agree to cite our paper "Quality Prediction for Super-Resolution Images with Low-Quality Reference" in any publications or presentations that utilize this dataset.

The Super-Resolution Combined Distortion(SRCD) dataset includes 100,000 distorted images generated by applying a modified degradation model to the original images of KADIS-700k. Additionally, to account for scenarios where images are distorted or enhanced by SR methods and scaling factors, we include four SR methods and two scaling factors. The dataset construction process considers three aspects: distortion type, SR method, and scaling factor.

Distortion Type

Low-quality broadcast content includes distortions from acquisition, processing, storage, and limitations of camera technology, even before applying SR. To reproduce this environment, common distortions found in broadcast content are randomly selected and applied. Additionally, various distortions are combined to generate combined distorted image. The types of distortions applied to generate the combined distorted image are as follow:

1. Blur: gaussian blur, motion blur

2. Noise: white noise, multi noise

3. Spatial distortion: jitter, pixelate

4. Compression: JPEG2000

5. Color brightness: color shift, brightness

Among the images, the image to which the SR will be applied performs downsampling and distortion is applied.

SR Method

Generally, recent SR methods tend to either amplify or reduce the distortions present in the images. Therefore, to account for both distortions and enhancements induced by SR, we include SR methods as factors influencing image characteristics. There are a total of four SR methods used, which can be categorized into methods that amplify distortions (Bicubic, SRGAN) and methods that remove distortions (BSRGAN, SwinIR).

Scaling Factor

Along with considering the SR method, the scaling factor is also taken into account as a factor influencing image characteristics. This consideration arises from the assumption that the performance of SR methods degrades as the scaling factor increases. Therefore, to measure quality changes due to the scaling factor, we include two commonly used scaling factors (×2, ×4) from the SR field.

Copyright (c) 2024 Intelligent Visual Media Laboratory (IVML)

All rights reserved.

The SRCD 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.