Online Supplement

Full-Reference Image Quality Assessment and the Role of Strategy: The Most Apparent Distortion

E. C. Larson and D. M. Chandler

Journal of Electronic Imaging, Special Section on Image Quality, 19 (1), March 2010

Supplementary Information

Notes on the Masking Model
MAD estimates the ability of each image region to mask distortions by comparing the RMS contrast of the distortions with the RMS contrast of the original image, in a blockwise fashion. In this note, we will illustrate this process on an example image turtle from the CSIQ image database.

Below are the original and distorted images (click thumbnails for full-sized versions). The distorted image contains Gaussian white noise.

Iorg Idst

The pixels of the input images are first converted to luminance values and then converted to relative perceived lightness values. Here, we assume sRGB display characteristics for the conversion from digital pixel values to physical luminances. The resulting images are shown below; these images have been rescaled to span the range [0, 255].

hat_Lorg hat_Ldst

Next, the contrast sensitivity function (CSF) originally described by Mannos and Sakrison with adjustments specified by Daly is applied to both images. This CSF is computed assuming that the maximum displayable spatial frequency is 16 c/deg in the horizontal or vertical direction. The resulting images are shown below; the error image has been scaled by a factor of 4 to promote visibility.

I'org I'err + 128

Next, we compare the (modified) RMS contrast of the orignal image to the RMS contrast of the distortions on a block-by-block basis (16x16 blocks with 75% overlap between neighboring blocks). The contrast of the distortions is computed relative to the mean of the original image. The contrast values of each block can be assembled into a "contrast map," which is shown below for the original and distorted images. To promote visibility, these contrast maps have been scaled by a factor of 7. (Note that these maps are still 8-bit images, so clipping has occured in the images below.)

Corg Cerr

Finally, to determine the visibility map, we compare each value of the original image's contrast map to the corresponding location in the error image's contrast map. This comparison is made using the following equation:

The resulting visibility map is shown below. Larger-valued pixels (whiter pixels) denote locations at which the distortions are progressively more visible. Zero-valued pixels (black pixels) denote locations at which the distortions are not visible.

Visibility Map (Eta)

Note that the model overestimates the masking ability of the region in the upper-left background and the region immediately above the turtle's head. This is due to the fact that original image's contrast is overestimated in this region (observe these same locations in the Corg map). This limitation can be overcome by using further sub-blocks when computing the modified standard deviation in the computation of the original image's contrast. Alternatively, a more proper masking model which takes into account image content (edge vs. texture vs. structure) may provide better results [1].

[1] D. M. Chandler, M. G. Gaubatz, and S. S. Hemami, "A Patch-Based Structural Masking Model with an Application to Compression," EURASIP Journal on Image and Video Processing Volume 2009 (2009), Article ID 649316, 22 pages doi:10.1155/2009/649316. [abstract]


Notes on Usage of the Code
Coming soon.


Links to the Other Databases Reported in the Paper
The following list provides access information for the databases reported in the paper.