Evaluation index
Image Quality Assessment (IQA) can be divided into subjective assessment and objective assessment from the method. Subjective evaluation is the quality score of the evaluation image and the reference image given by human eyes. Objective evaluation is to use mathematical models to give quantitative values, and image processing technology can be used to generate a batch of distorted images, which is easy to operate and has become the focus of IQA research. Commonly used objective evaluation indicators are PSNR and SSIM.
Peak Signal to Noise Ratio (Peak Signal to Noise Ratio, PSNR)
It is usually used to evaluate the quality of an image after compression compared with the original image. The higher the PSNR, the smaller the distortion after compression. It can be calculated by means of Mean Square Error (MSE). The calculation formula is:
Structural similarity index (SSIM)
It is an indicator used to measure the similarity of two images. When one of the two images is an undistorted image and the other is a distorted image, the structural similarity of the two images can be regarded as a quality indicator of the distorted image. Structural similarity Compared with peak signal-to-noise ratio, the structural similarity index is more in line with the judgment of human eyes on image quality in the measurement of image quality.
The formula is as follows, C1, C2 are constants, μ, σ are mean and standard deviation.
General Image Quality Assessment Dataset
The widely recognized datasets are: LIVE, TID2008, TID2013, CSIQ, IVC and MICT (Toyama). Each dataset gives the mean subjective score (MOS) of the images.
If the official website download is slow, you should be able to find some domestic fast download addresses.
1) LIVE (Laboratory for image & video engineering)
http://live.ece.utexas.edu/index.php
LIVE is jointly established by the Department of Electrical and Computer Engineering and the Department of Psychology at the University of Texas at Austin. most widely. The Release 2 version contains 29 reference images and 779 distorted images, including 175 JPEG2000 distorted images, 169 JPEG distorted images, 145 white noise distorted images, 145 Gaussian blur distorted images, and 145 fast Rayleigh attenuated distorted images. The DMOS value of the database is obtained from about 25,000 data statistics given by 161 observers, and the value range of DMOS is [0,100].
2) TID2008 (Tampere image database)
http://www.ponomarenko.info/tid2008.htm
TID2008 was established by the N504 Department of Signal Reception, Transmission and Processing of Ukraine National Aerospace University, including 25 reference images and 1700 distorted images . There are 17 types of distortion including: additive Gaussian noise, additive noise where color components are stronger than lighting components, spatial position-dependent noise, mask noise, high-frequency noise, impulse noise, quantization noise, Gaussian blur, image noise, JPEG compression , JPEG2000 compression, JPEG transmission errors, JPEG2000 transmission errors, non-eccentric noise, local block distortions of different intensities, intensity mean shifts, and contrast changes. The DMOS value of the database is obtained from 256,428 data statistics given by 838 observers, and the MOS value range is [0,9].
3) TID2013 (Tampere image database)
http://www.ponomarenko.info/tid2013.htm
TID2013 is an enhanced version of TID2008, including 25 reference images and 3000 distorted images. There are 24 kinds of distortion types, including: changing color saturation, multiple Gaussian noise, comfort noise, lossy compression, color image quantization, color difference and sparse sampling. The DMOS value of the database is obtained from 524,340 data statistics given by 971 observers, and the value range of MOS is [0,9]. Because the database has many types of distortion, the database is more abundant, and it is a color distortion database, so more and more algorithms include this database in comparative experiments.
4) CSIQ (Categorical subjective image quality)
http://vision.okstate.edu/csiq
CSIQ was established by the School of Electrical and Computer Engineering at Oklahoma State University in the United States, including 30 reference images, 866 distorted images, and distortion types Including JPEG compression, JPEG2000 compression, overall contrast reduction, additive Gaussian pink noise, additive Gaussian white noise and Gaussian blur. The DMOS value of the database is obtained from about 5000 data statistics given by 25 observers, and the value range of DMOS is [0,1].
5) IVC
http://www2.irccyn.ec-nantes.fr/ivcdb/
IVC was established by the Central Polytechnic University of Nantes, France, including 10 reference images and 235 distorted images. Distortion types include JPEG compression, JPEG2000 compression, LAR encoding, and blurring. The MOS value of the database is obtained from statistics given by 15 observers, and the value range of MOS is [0,5].
6) MICT
http://mict.eng.u-toyama.ac.jp/mictdb.html
MICT was founded by Toyama University, including 14 reference images and 168 distorted images. Distortion types include JPEG compression and JPEG2000 compression. The MOS value of the database is obtained from statistics given by 16 observers, and the value range of MOS is [1,5].
7) A57
http://foulard.ece.cornell.edu/dmc27/vsnr/vsnr.html
A57 was created by Cornell University, including 3 reference images and 54 distorted images. Distortion types include: a) For the quantization on the 5 LH subbands after the discrete Xiaobo transform of the image, the uniform step size is used for quantization, and the mean square error of the contrast of the distorted image is equal. b) Additive white Gaussian noise c) JPEG compression d) JPEG2000 compression without visual frequency weighting e) JPEG2000 compression with a quantization algorithm based on dynamic contrast f) Gaussian blur The DMOS values of the database are given by 7 observers Data statistics It is obtained that the value range of MOS is [0,1].
8) WIQ (Wireless imaging quality)
http://www.bth.se/tek/rcg.nsf/pages/wiq-db
WIQ is a researcher from Blekinge Institute of Technology in Sweden and Siguna Del Mar University in Indonesia Established in cooperation, including 7 reference images and 80 distorted images. There are 5 types of distortion, including: "flat" distribution, JPEG compression, JPEG2000 compression, JPEG200+DCQ compression, Gaussian blur and Gaussian white noise. The DMOS value of the database is given by 60 observers
as a database of image quality evaluation_lanmengyiyu's blog-CSDN blog_image quality evaluation database
Super Resolution Image Quality Assessment Dataset
The evaluation data sets of the super-resolution algorithm are: Set5, Set14, BSD100, Urban100, there will be some classic images of different resolution sizes, all on this website: eugenesiow (Eugene Siow )
Some models and performance situations:
GitHub - eugenesiow/super-image: State-of-the-art image super resolution models for PyTorch.
