Image Quality Measures Software




Abstract

 

The Image Quality Toolbox is a Matlab Toolbox for full reference color (and also achromatic) image quality assessment based on divisive normalization models in DCT and wavelet domains.
The general idea to assess the perceptual distance between two images is to compute the q-norm Euclidean distance in the image representation at the V1 visual cortex, as suggested in [Teo&Heeger, IEEE ICIP 1994]. This markedly differs from the Mean Square Error (2-norm Euclidean measure in the spatial domain), as shown in [Pons99, Epifanio03]
 

These ideas have been implemented in the wavelet domain in the new code associated to the paper [Laparra10a]. Results using the wavelet based measure outperform SSIM and VIF and are intuitively interpretable in a linear way.

  • DCT distance: The original ideas for the DCT-based distance presented here date back to the late 90's [Malo97, Pons99]. The use of the V1 representation with different summation norms gives rise to the Maximum Perceptual Error (MPE) concept, as used in a number of our DCT image/video coding algorithms [Malo99 ,  Malo00 , Malo01 , Epifanio03 , Gomez05 , Malo06 , Camps08].
    The current version of the algorithm uses parameters to reproduce achromatic contrast incremental thresholds. Extension to color channels was done by using the same parameters in the non-linearity but using the Mullen chromatic CSFs in the blue-yellow (U) and red-green (V) channels [Mullen, J.Physiol.1985] instead of the achromatic band-pass sensitivity. Achromatic and chromatic CSFs have been relatively scaled using an appropriate chromatic contrast measure. No additional fit to any subjectivelly rated image data base has been done. Next version of the toolbox will include optimized parameters.
  • Wavelet distance: The wavelet-based distance measure only differs from the DCT one in the following aspects:

    1. The initial linear transform is an orthogonal QMF wavelet instead of block DCT

    2. As a result, the kernel in the divisive normalization may also include spatial masking (beyond the frequency masking considered in the DCT case).

    3. Parameters were optimized to maximize the correlation among the predictions of the model and the Mean Opinion Score (MOS) on a subjectively rated data base: the JPEG2000 subset of the LIVE database (http://live.ece.utexas.edu/research/quality/).




Related Papers


  • DCT distance:
Image quality metric based on multidimensional contrast perception models
A.M. Pons, J. Malo, J.M. Artigas and P.Capilla
Displays Journal, Vol. 20, pp. 93-110. (1999)
pdf

Linear Transform for Simultaneous Diagonalization of Covariance and Perceptual Metric Matrix in Image Coding
I. Epifanio, J.Gutiérrez, J.Malo
Pattern Recognition.  Vol.36, pp. 1799-1811 (2003) pdf
 
Subjetive image fidelity metric-based on bit allocation of the human visual system in the DCT domain.
J. Malo, A.M. Pons and J.M. Artigas
Image and Vision Computing, Vol. 15, 7, pp 535-548 (1997)
pdf

Comparison of perceptually uniform quantization with average error minimization in image tranform coding
J.Malo, F.Ferri, J.Albert, J.Soret
Electronics Letters, Vol. 35, 13, pp. 1067-1068 (1999)  pdf

The role of perceptual contrast non-linearities in image transform quantization
J. Malo, F. Ferri, J. Albert, J. Soret and J.M. Artigas
Image & Vision Computing, Vol. 18, 3, pp. 233-246 (2000)
pdf

Perceptual feed-back in multigrid motion estimation using an improved DCT quantization
J.Malo, J.Gutierrez, I.Epifanio, F.Ferri, and J.M.Artigas
IEEE Trans. Im. Proc. Vol.10, 10, pp. 1411-1427 (2001) pdf

Perceptual Adaptive Insensitivity for Support Vector Machine Image Coding
G. Gómez, G. Camps, J. Gutiérrez, and J. Malo, 
IEEE Trans. Neural Networks
Vol. 16, 6, pp 1574-1581 (2005) pdf

Non-linear Image Representation for Efficient Perceptual Coding
J. Malo, I. Epifanio, R. Navarro and E.P. Simoncelli, 
IEEE Trans. Im. Proc.
Vol. 15, 1, pp 68-80 (2006) pdf

On the Suitable Domain for SVM Training in Image Coding
G. Camps, J. Gutiérrez, G. Gómez and J. Malo  
Journal of Machine Learning Research
, Vol. 9, pp 49-66 (2008) pdf



  • Wavelet distance:
Divisive Normalization Image Quality Metric Revisited.
V. Laparra, J. Muñoz and J. Malo.
JOSA A, 27(4): 852-864 (2010)
.
Also selected by the OSA for the Virtual Journal for Biomedical Optics 5(8), 2010.
pdf




Download Code
 

DCT distance:


vistaqualitytools.zip mat


Wavelet distance:

div_norm_metric.zipmat