CoChat

Blind Prediction of Natural Video Quality

Saad, Michele A., Bovik, Alan C., Charrier, Christophe · 2014 · IEEE Transactions on Image Processing

The DOI matches its Crossref record; 60 of the 60 references it cites were found in a registry or an index; none of the 22 sources its claims cite is checked for support yet.

Abstract

We propose a blind (no reference or NR) video quality evaluation model that is nondistortion specific. The approach relies on a spatio-temporal model of video scenes in the discrete cosine transform domain, and on a model that characterizes the type of motion occurring in the scenes, to predict video quality. We use the models to define video statistics and perceptual features that are the basis of a video quality assessment (VQA) algorithm that does not require the presence of a pristine video to compare against in order to predict a perceptual quality score. The contributions of this paper are threefold. 1) We propose a spatio-temporal natural scene statistics (NSS) model for videos. 2) We propose a motion model that quantifies motion coherency in video scenes. 3) We show that the proposed NSS and motion coherency models are appropriate for quality assessment of videos, and we utilize them to design a blind VQA algorithm that correlates highly with human judgments of quality. The proposed algorithm, called video BLIINDS, is tested on the LIVE VQA database and on the EPFL-PoliMi video database and shown to perform close to the level of top performing reduced and full reference VQA algorithms.

Overview

CoChat Summary

TL;DR

The paper proposes Video BLIINDS, a blind/no-reference video quality assessment model based on spatio-temporal natural video statistics in the DCT domain plus motion coherency and global motion features. It predicts perceptual video quality without a pristine reference and performs close to top reduced- and full-reference VQA algorithms on the LIVE VQA and EPFL-PoliMi databases.

Key findings

  • Video BLIINDS achieved median SROCC/LCC on the LIVE VQA database of 0.869/0.924 for MPEG-2, 0.839/0.893 for H.264, 0.815/0.951 for wireless, 0.779/0.946 for IP, and 0.759/0.881 when all distortions were mixed.
  • It clearly outperformed the blind NIQE index, whose mixed-distortion SROCC/LCC was 0.151/0.317, and also outperformed full-reference PSNR and SSIM; it did not quite reach top full-reference methods such as ST-MAD and MOVIE but was statistically similar to reduced-reference VQM on LIVE.
  • In feature-isolation tests, NVS shape-parameter ratios gave the highest prediction SROCC at 0.587, followed by the DC feature at 0.486, spatial naturalness at 0.324, global motion at 0.220, and coherency at 0.204.
  • On EPFL-PoliMi, Video BLIINDS achieved median SROCC/LCC of 0.807/0.752 versus NIQE’s 0.555/0.124; on homogeneous EPFL content, its median SROCC rose to 0.989 versus NIQE’s 0.6978.
  • Two H.264 applications were demonstrated: bit-rate selection achieved median SROCC 0.954 with median MSE 0.374 MB, and bit-rate prediction achieved median SROCC 0.860 with median MSE 0.471 MB.

Methods

  • The model computes frame differences, applies a 5×5 DCT to local patches, fits a generalized Gaussian density to each DCT frequency coefficient, and forms matrices of shape parameters γ per frame difference. It derives low/mid/high spectral ratios from these γ matrices, plus a temporal derivative of mean DC coefficients, motion coherency from a 2D structure tensor on motion vectors, a global motion measure, and NIQE-based spatial naturalness features.
  • Features are logarithmically transformed and temporally pooled over 10-second intervals, then used to train a linear kernel support vector regressor for quality prediction.
  • Evaluation used the LIVE VQA database (160 videos from 10 reference videos, four distortion types) with content-independent 80%/20% train/test splits over all combinations, reporting SROCC and PLCC; EPFL-PoliMi was also used, with leave-one-out training/testing because it contains only six 4-CIF reference videos.

Limitations

  • The paper states that no existing non-distortion-specific blind VQA approaches were available for direct comparison, that blind VQA algorithms trained on a database are applicable to the distortions present in training, and that more complete motion and natural video statistics models are still needed.
  • The paper also states that temporal distortion modeling increases computational complexity, that large video corpora are needed for validation, and that its aim is to reach full-reference-level correlations.
  • Evident from the text: Video BLIINDS still requires training on human quality scores, does not match top full-reference VQA performance on the mixed LIVE set, and the bit-rate applications assume a particular H.264 encoder parameter configuration.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tip.2014.2299154

Cites

60 of 60 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Left: frame from pristine video. Right: frame from distorted video.
Fig. 1 Left: frame from pristine video. Right: frame from distorted video.From the PDF
Image 2
From the PDF
Fig. 2 Empirical probability distribution of frame-difference DCT coefficients of pristine and distorted videos. Dashed line: pristine video. Solid line: distorted video.
Fig. 2 Empirical probability distribution of frame-difference DCT coefficients of pristine and distorted videos. Dashed line: pristine video. Solid line: distorted video.From the PDF
Fig. 3 Empirical probability distribution of frame-difference DCT coefficients of pristine and distorted videos for 4 distortions (MPEG-2, H.264, IP, and wireless distortions). Dashed line: pristine video. Solid line: distorted video.
Fig. 3 Empirical probability distribution of frame-difference DCT coefficients of pristine and distorted videos for 4 distortions (MPEG-2, H.264, IP, and wireless distortions). Dashed line: pristine video. Solid line: distorted video.From the PDF
Fig. 4 Plot of video quality parameter γ over time for three videos. Stars (top): pristine video, DMOS = 0. Thin black (middle): medium quality video, DMOS = 56 . 1328. Gray (bottom): low quality video, DMOS = 72 . 1356.
Fig. 4 Plot of video quality parameter γ over time for three videos. Stars (top): pristine video, DMOS = 0. Thin black (middle): medium quality video, DMOS = 56 . 1328. Gray (bottom): low quality video, DMOS = 72 . 1356.From the PDF
Fig. 5 Blind VQA framework.
Fig. 5 Blind VQA framework.From the PDF
Fig. 6 Empirical distribution of DCT coefficients at 5 different frequencies from an n × n DCT decomposition of a frame-difference.
Fig. 6 Empirical distribution of DCT coefficients at 5 different frequencies from an n × n DCT decomposition of a frame-difference.From the PDF
Fig. 7 Generalized Gaussian density plots for different values of the shape parameter γ .
Fig. 7 Generalized Gaussian density plots for different values of the shape parameter γ .From the PDF
Fig. 8 n × n matrix of the shape parameter γ values is obtained for each frame-difference, by fitting a GGD model to the histogram of each frequency in the n × n DCT block over all blocks in a frame-difference.
Fig. 8 n × n matrix of the shape parameter γ values is obtained for each frame-difference, by fitting a GGD model to the histogram of each frequency in the n × n DCT block over all blocks in a frame-difference.From the PDF
Image 10
From the PDF
Image 11
From the PDF
Image 12
From the PDF
Distorted Ice-Skating Snapshot
Distorted Ice-Skating SnapshotFrom the PDF
Fig. 12 Plot of the temporal derivative of mean DC coefficients for a pristine and a distorted video.
Fig. 12 Plot of the temporal derivative of mean DC coefficients for a pristine and a distorted video.From the PDF
Fig. 13 The spatio-temporal features (DC feature, coherency measure, global motion measure, and shape-parameter spectral ratios are logarithmically transformed before being used as features for quality prediction by the SVR).
Fig. 13 The spatio-temporal features (DC feature, coherency measure, global motion measure, and shape-parameter spectral ratios are logarithmically transformed before being used as features for quality prediction by the SVR).From the PDF
Fig. 14 Plot of median SROCC distribution for PSNR, SSIM, VQM, NIQE, and Video BLIINDS.
Fig. 14 Plot of median SROCC distribution for PSNR, SSIM, VQM, NIQE, and Video BLIINDS.From the PDF
Fig. 17 Application 1: Histogram of SROCC between predicted and actual bit rates over 100 iterations of train/test splits.
Fig. 17 Application 1: Histogram of SROCC between predicted and actual bit rates over 100 iterations of train/test splits.From the PDF
Fig. 18 Application 1: Scatter plot of predicted versus actual bit rates.
Fig. 18 Application 1: Scatter plot of predicted versus actual bit rates.From the PDF
Fig. 19 Application 1: Scatter plot of selected versus actual bit rates.
Fig. 19 Application 1: Scatter plot of selected versus actual bit rates.From the PDF
Fig. 20 Application 1: Histogram of SROCC between selected and actual bit rates over 100 iterations of train/test splits.
Fig. 20 Application 1: Histogram of SROCC between selected and actual bit rates over 100 iterations of train/test splits.From the PDF
Image 21
From the PDF
Image 22
From the PDF
Image 23
From the PDF

Verification

Claims the paper makes

0 Valid · 2 Not valid · 20 Not checked · 58 Own results and interpretations

44 more claims

Create a free account to see all 80 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 17
  • Probably not a substantial claim. 17
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 60 references

Research works printed but not linked: 2

  1. blake, 2006, Perception
  2. 2007, H 264/MPEG-4 AVC Reference Software Manual

Cited by

Items 301–400 of 471

Metadata

Authors
Saad, M. A., Bovik, A. C., & Charrier, C.
Affiliations
  • The University of Texas at Austin (US)
  • Université de Caen Normandie (FR)
Venue
IEEE Transactions on Image Processing
Published
2014-03
Volume
23
Issue
3
Pages
1352-1365
ISSN
  • 1057-7149 (ISSN-L)
  • 1941-0042
OpenAlex
W2048042940
Cited by
471
References
57
Citations by year
2026: 26 · 2025: 47 · 2024: 57 · 2023: 59 · 2022: 48 · 2021: 57 · 2020: 19 · 2019: 30 · 2018: 36 · 2017: 40 · 2016: 21 · 2015: 17 · 2014: 14
FWCI
27.40
Type
journal article
Language
en
Open access
closed
Locations
  • IEEE Transactions on Image Processing (publishedVersion)
  • PubMed (publishedVersion)
  • HAL (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Video Quality Assessment
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Retrieval and Classification Techniques
  • Physical Sciences › Computer Science › Signal Processing › Video Coding and Compression Technologies
Keywords
blind video quality assessment, natural scene statistics, spatio-temporal modeling, discrete cosine transform, motion modeling, perceptual quality
MeSH terms
Algorithms, Image Enhancement, Image Interpretation, Computer-Assisted, Motion, Pattern Recognition, Automated, Photography, Sensitivity and Specificity, Video Recording, Reproducibility of Results, Artifacts

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.