KADID-10k Database

A large-scale artificially distorted
image quality assessment database

Current artificially distorted image quality assessment (IQA) databases are small in size and limited in content. Larger IQA databases that are diverse in content could benefit the development of deep learning based IQA methods.

For this purpose, we created two datasets, the Konstanz artificially distorted image quality database (KADID-10k) and the Konstanz artificially distorted image quality set (KADIS-700k). The former contains 81 pristine images, each degraded by 25 distortions in 5 levels. The latter has 140,000 pristine images, with 5 degraded versions each, where the distortion was chosen randomly. Through the use of crowdsourcing, we conducted a subjective IQA study on KADID-10k and obtained 30 degradation category ratings (DCRs) per image.

Cite us

KADID-10k is freely available to the research community. If you use our databases and/or distortion code in your research, please cite it as follows:

@inproceedings{kadid10k,
title={KADID-10k: A Large-scale Artificially 
Distorted IQA Database},
author={Lin, Hanhe and Hosu, Vlad and Saupe, Dietmar},
booktitle={2019 Tenth International Conference on 
Quality of Multimedia Experience (QoMEX)},
pages={1--3},
year={2019},
organization={IEEE}}

@article{deepfl-iqa,
title={DeepFL-IQA: Weak Supervision for Deep IQA 
Feature Learning},
author={Lin, Hanhe and Hosu, Vlad and Saupe, Dietmar},
journal={arXiv preprint arXiv:2001.08113},
year={2020}}

Alternative download

Downloads

KADID-10k (.ZIP)

~ 3.1 GB

The KADID-10k database contains:

"image" folder: contains 81 reference images and 10,125 distorted images (81 reference images x 25 types of distortion x 5 levels of distortions). All images are saved as PNG format. The name format for distorted image is Ixx_yy_zz.png, where xx is the image id, yy is the type of distortion, zz is the level of distortion. For more information about distortion types, please see below.
"dmos.csv" file: contains the differential mean opinion score (DMOS) as well as variance for each distorted image. The range of DMOS is [1 5], high value of DMOS corresponds to higher visual quality of image.

KADID-10K raw data

~68.5 MB

kadis-700K (.zip)

download mirror

~ 44.6 GB

The KADIS-700 dataset contains:

"ref_imgs" folder: contains 140,000 reference images.
"dist_imgs" folder: empty, used for saving 700,000 distorted images.
"code_imdistort" folder: contains the MATLAB code for all 25 types of distortions
"kadis700k_ref_img.csv" file: contains the information for generating 700,000 distorted images.
"generate_kadis700.m" file: script to generate 700,000 distorted image.

matlab code fOR DISTORTED IMAGE GENERATION

The 81 pristine images in KADID-10k

Interface for subjective IQA study

Histogram of DMOS

Distortion types

Blurs
- # 01 Gaussian blur: filter with a variable Gaussian kernel
- # 02 Lens blur: filter with a circular kernel
- # 03 Motion blur: filter with a line kernel
Color distortions
- # 04 Color diffusion: Gaussian blur the color channels (a and b) in the Lab color-space
- # 05 Color shift: randomly translate the green channel, and blend it into the original image masked by a gray level map: the normalized gradient magnitude of the original image
- # 06 Color quantization: convert to indexed image using minimum variance quantization and dithering with 8 to 64 colors
- # 07 Color saturation 1: multiply the saturation channel in the HSV color-space by a factor
- # 08 Color saturation 2: multiply the color channels in the Lab colorspace by a factor
Compression
- # 09 JPEG2000: standard compression
- # 10 JPEG: standard compression
Noise
- # 11 White noise: add Gaussian white noise to the RGB image
- # 12 White noise in color component: add Gaussian white noise to the YCbCr converted image (both to the luminance ‘Y‘ and the color channels ‘Cb‘ and ‘Cr‘)
- # 13 Impulse noise: add salt and pepper noise to the RGB image
- # 14 Multiplicative noise: add speckle noise to the RGB image
- # 15 Denoise: add Gaussian white noise to RGB image, and then apply a denoising DnCNN to each channel separately
Brightness change
- # 16 Brighten: non-linearly adjust the luminance channel keeping extreme values fixed, and increasing others
- # 17 Darken: similar to brighten, but decrease other values
- # 18 Mean shift: add constant to all values in image, and truncate to original value range
Spatial distortions
- # 19 Jitter: randomly scatter image data by warping each pixel with random small offsets (bicubic interpolation)
- # 20 Non-eccentricity patch: randomly offset small patches in the image to nearby locations
- # 21 Pixelate: downsize image and upsize it back to the original size using nearest-neighbor interpolation in each case
- # 22 Quantization: quantize image values using N thresholds obtained using Otsus method
- # 23 Color block: insert homogeneous random colored blocks at random locations in the image
Sharpness and contrast
- # 24 High sharpen: over-sharpen image using unsharp masking
- # 25 Contrast change: non-linearly change RGB values using a Sigmoid-type adjustment curve