1. A Review in Quality Measures for Halftoned Images
Student Per-Erik Axelson

2. Ph.D. Course in Digital Halftoning
Image Quality
Subjective Quality
Subjective test (MOS): Best way so far to assess and judge image quality
This method is to inconvenient, slow and expensive for practical usage
Objective Quality Measures
The goal of objective image quality assessment research is to supply quality metrics that can predict perceived image quality automatically
29/11/2018
Ph.D. Course in Digital Halftoning

3. Objective Image Quality
Image quality paradigm
Original Image
Reproduced Image
Most important demand
In General, we want binary image to match continuous-tone original as closely as possible
29/11/2018
Ph.D. Course in Digital Halftoning

4. Objective Image Quality
A number of demands
The method should be able to evaluate all kinds of halftones and provide a meaningful comparison across different types of image distortions
The method should return measures for several aspects of quality that are well correlated with results from subjective tests
The method should be easy to calculate and have low computational complexity
29/11/2018
Ph.D. Course in Digital Halftoning

5. Objective Image Quality
Useful applications
Be used to monitor image quality for quality control systems
Be employed to benchmark halftoning algorithms
Be embedded into an image processing system to optimize the algorithms and the parameter settings
29/11/2018
Ph.D. Course in Digital Halftoning

6. Objective Image Quality
Two classes of objective quality assessment
1. Mathematically defined measures
Methods based on the Mean Square Error (MSE)
2. Models of the human visual system (HVS)
Methods using the contrast sensitivity function (CSF)
29/11/2018
Ph.D. Course in Digital Halftoning

7. Image Quality Error Metrics
Function
Derive measure from the point-wise difference between original and the binary halftone
In general, using a fixed threshold at midpoint
Definition
29/11/2018
Ph.D. Course in Digital Halftoning

8. Image Quality Error Metrics
Correlation between MSE, SNR or PSNR and visual quality is known to be poor
Treats all errors with an equal weight
White Noise
SNR = 10 dB
PSNR = 15,7 dB
High-pass Noise
SNR = 10 dB
PSNR = 15.7
29/11/2018
Ph.D. Course in Digital Halftoning

9. Ph.D. Course in Digital Halftoning
Human Visual System
Complicated Non-linear and spatially varying
Assuming linearity and spatial invariance
The human perception system do not have equal response to all spatial frequencies
As the spatial frequencies become higher and higher, our ability to perceive the pattern will be lower and lower
It turns out that our ability to perceive very low frequency patterns also decreases as the frequency decreases
These characteristics can be captured using a contrast sensitivity function (CSF)
29/11/2018
Ph.D. Course in Digital Halftoning

10. Ph.D. Course in Digital Halftoning
Human Visual Response
Sensitivity depends on angular frequency subtended at eye
Compute angular frequency from image size (pixels), printed image size (mm), viewing distance (mm)
Object image by the eye
Visual angle:
At Nyquist frequency:
29/11/2018
Ph.D. Course in Digital Halftoning

11. Contrast Sensitivity Function
Band-pass model [Mannos & Sakrison 1974]
Modified to low-pass [Mitsa & Varkur 1993]
29/11/2018
Ph.D. Course in Digital Halftoning

12. Contrast Sensitivity Function
Angular dependence in CSF [Sullivan, Miller & Pios 1993]
Mild-drop in visual sensitivity in diagonal directions
The decreased sensitivity along the diagonals and the flattening at low angular frequencies are visible
29/11/2018
Ph.D. Course in Digital Halftoning

13. Ph.D. Course in Digital Halftoning
Weighted SNR Metric
Weighted SNR by CSF
WSNR measures appropriate when noise is additive and signal independent
Where X(u,v), Y(u,v) and C(u,v) represent the DFT of the input image, output image and CSF, respectively
29/11/2018
Ph.D. Course in Digital Halftoning

14. Ph.D. Course in Digital Halftoning
WSNR
To find WSNR
Generate unsharpened halftone using modified error diffusion [Eschbach & Knox 1991]
Compute WSNR of unsharpened halftone relative to original image
29/11/2018
Ph.D. Course in Digital Halftoning

15. A Universal Image Quality Index
Main Features
New Philosophy: switch from error measurement to structural distortion measurement
Mathematically defined and no HVS model is explicitly employed
Universal: Applicable on various image-processing applications and provide a meaningful comparison across different types of image distortions
Easy to apply on images
Low computational complexity
29/11/2018
Ph.D. Course in Digital Halftoning

16. A Universal Image Quality Index
Application to Images
Compare difference between the original and the binary image
Measure statistical features locally and then combine them together
Sliding window (size 8  8) approach in local region, leading to a quality map
The index value is the average of the quality map
29/11/2018
Ph.D. Course in Digital Halftoning

17. A Universal Image Quality Index
Definition
1 2 3
Q : Dynamic range [-1, 1]
Combination of three factors
1. Loss of Correlation (Linear correlation between x and y)
2. Luminance Distortion (Mean luminance between x and y)
3. Contrast Distortion (Variance contrast (signal) between x and y)
29/11/2018
Ph.D. Course in Digital Halftoning