Results 1 to 3 of 3

Thread: Mean Opinion Score(MOS) for Image compression

  1. #1
    Member
    Join Date
    Aug 2017
    Location
    Mauritius
    Posts
    59
    Thanks
    67
    Thanked 22 Times in 16 Posts

    Post Mean Opinion Score(MOS) for Image compression

    What are the current MOS data sets that are used in image compression domain ?.The methodology of the MOS tests is also very important as there are a lot of variables at play here.In my humble opinion factors influencing the experiment are as follows:


    1)Equipment and environment
    Screen used and the distance of the test subject , the lightning conditions as well as how well trained the test subjects are to detect compression artifacts in images.


    2)How diverse are the images in the dataset ?


    There are a lots of different types of images in the world namely photography(aerial , portrait , scenic ), black / white , medical images etc.The image dataset must represent the diversity in the world.


    3)Defining the relationship - (bpp -> MOS score) for each image.


    There are different standards for quality depending on the use cases such as psychotically lossless , "Good" , "Bad" )


    4) Source images


    I believe each source image should have the following variants each with their own MOS score when applicable


    - The RAW image (the raw image is the source of truth and contain minimal bias).Moreover raw images can be used to generate other images with different settings (i.e light balance, special effects etc...)
    - The JPEG image generated by the camera.
    - The RAW image enhance by a photographer saved in a lossless format.


    An efficient subjective image quality assessment can help to build objective quality metrics of the future , image formats as well as help research as a whole given the high costs associated with subjective image assessment.


    I wonder what are the views of the compression experts , what factors i have missed out and how it can improved ?

  2. #2
    Member
    Join Date
    Apr 2012
    Location
    Stuttgart
    Posts
    437
    Thanks
    1
    Thanked 96 Times in 57 Posts
    Quote Originally Posted by khavish View Post
    I wonder what are the views of the compression experts , what factors i have missed out and how it can improved ?
    Most importantly, the test protocol. So is this "side-by-side" (double-stimulus), single image (single-stimulus), with or without "with hidden reference"? How long is the observation time? How long lasts a session? How many observers have been selected for the test? What is their average age? There are also "continuous scale" protocols where observers select, rather from a 5-point scale, assess quality from a 100-point scale.

  3. #3
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    667
    Thanks
    204
    Thanked 241 Times in 146 Posts
    Quote Originally Posted by khavish View Post
    2)How diverse are the images in the dataset ?
    Khavish, thank you for asking. Here are some ideas...

    1. In normal images that are compressed the bulk of the information (80-97 %) are in the relatively high frequency bands. In jpeg terms this information is in the coefficients [1..63], i.e., dc is excluded.

    2. Highest spatial frequency human color vision is bichromatic as there are no S receptors in fovea. Bichromatic is crossed into L - M and L + M components. Because of this the images should contain HF information that specifically is designed to excite L - M, L + M or S receptors. That way the images will expose how well the compression system deals with the bichromacity. Often YUV420 is used, but it assumes purely monochromatic HF vision and consequently visible artefacts are created.

    3. The need for gamma correction emerges from opsin dynamics (a differential equation how opsins travel between excited and ground state.) The opsin dynamics do not depend on the R, G and B components, so traditional gamma correction is a common and coarse mistake that does not quite work. Because of this one needs to have different color situations modeled as a base color, and then have examples of high frequency information in different color contexts. At minimum every corner of the RGB space should be included and a few points in the middle. None of the image corpora (except my own :-P) that I have seen consciously satisfies this basic requirement. Some have a strong emphasis on red backgrounds because empirically people have seen them reproduced badly.

    4. The high frequency information should contain visual features that match basic integration primitives on retina. This includes DoG and lines. Thin lines are often removed very early by image compressors, even when they can be prominent features.

    5. It should contain faint directional textures that are in the proximity of a high intensity edge or ridge that has an orthogonal directivity. This can confuse many advanced compressors that otherwise do a decent job.

    6. It should contain textures that have a low amplitude and high amplitude component.

    7. It should contain periodic textures to see how well phase is preserved from one block to next -- the periodicity can indicate quality of cloth and messing it up changes semantics.

    8. It should contain images with different levels and structure of acquisition noise. It should contain images practically without noise, too.

    9. It should contain images built with different demosaicking algorithms as well as images without demosaicking artefacts (downsampled).

    10. It should contain features that are oblique -- those are often 7x more difficult for integral transforms than just straight images.

    11. An ideal corpus should represent actual use.

    12. The corpus should contain images with text in the native language (or with a font that is compatible with the native language) to the eventual human rater population.

    13. When images are processed (resampled, raw->jpeg processed) ideally at least some of the images in the corpus would follow the best practices of professional photographers, i.e., includes unsharp masking, subsampling, possibly HF color suppression. Eventually these best practices will trickle down to consumers and would be great if the format is evaluated to be working well for these tricks.

    14. Once the technical diversity characteristics are filled, there is room for artistic and creative diversity, too. The images should be a pleasure to look at

Similar Threads

  1. UCI Image Compression
    By maadjordan in forum Data Compression
    Replies: 5
    Last Post: 19th August 2017, 23:15
  2. Your opinion on matrixview.com?
    By necros in forum Data Compression
    Replies: 3
    Last Post: 20th March 2017, 22:13
  3. What`s your opinion on browser cache in RAMdisk?
    By necros in forum The Off-Topic Lounge
    Replies: 6
    Last Post: 14th January 2017, 13:12
  4. Image compression for microcontroller
    By branislav in forum Data Compression
    Replies: 6
    Last Post: 16th November 2015, 23:24
  5. 3d image compression
    By m^2 in forum Data Compression
    Replies: 5
    Last Post: 9th July 2012, 08:52

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •