 |
Image Quality Assessment (IQA) 1.1.2
A fast, accurate, and reliable C library for measuring image quality.
|
|
Image Quality Assessment (IQA) 1.1.2
A fast, accurate, and reliable C library for measuring image quality.
|
All of the algorithms described here are called full-reference algorithms. This means they required the original undistorted image to compare the distorted image against.
Mean Squared Error is the average squared difference between a reference image and a distorted image. It is computed pixel-by-pixel by adding up the squared differences of all the pixels and dividing by the total pixel count.
For images A = {a1 .. aM} and B = {b1 .. bM}, where M is the number of pixels:
The squaring of the differences dampens small differences between the 2 pixels but penalizes large ones.
More info: http://en.wikipedia.org/wiki/Mean_squared_error
Peak Signal-to-Noise Ratio is the ratio between the reference signal and the distortion signal in an image, given in decibels. The higher the PSNR, the closer the distorted image is to the original. In general, a higher PSNR value should correlate to a higher quality image, but tests have shown that this isn't always the case. However, PSNR is a popular quality metric because it's easy and fast to calculate while still giving okay results.
For images A = {a1 .. aM}, B = {b1 .. bM}, and MAX equal to the maximum possible pixel value (2^8 - 1 = 255 for 8-bit images):
More info: http://en.wikipedia.org/wiki/PSNR
Structural SIMilarity is based on the idea that the human visual system is highly adapted to process structural information, and the algorithm attepts to measure the change in this information between and reference and distorted image. Based on numberous tests, SSIM does a much better job at quantifying subjective image quality than MSE or PSNR.
At a high level, SSIM attempts to measure the change in luminance, contrast, and structure in an image. Luminance is modeled as average pixel intensity, constrast by the variance between the reference and distorted image, and structure by the cross-correlation between the 2 images. The resulting values are combined (using exponents referred to as alpha, beta, and gamma) and averaged to generate a final SSIM index value.
The original paper defined 2 methods for calculating each local SSIM value: an 8x8 linear or 11x11 circular Gaussian sliding window. IQA defaults to using the Gaussian window that the paper suggests for best results. However, the window type, stabilization constants, and exponents can all be set adjusted by the application.
Original paper: https://ece.uwaterloo.ca/~z70wang/publications/ssim.html
Here's an interesting article by the authors discussing the limitations of MSE and PSNR as compared to SSIM: https://ece.uwaterloo.ca/~z70wang/publications/SPM09.pdf
Multi-Scale Structural SIMilarity is layered on SSIM. The algorithm calculates multiple SSIM values at multiple image scales (resolutions). By running the algorithm at different scales, the quality of the image is evaluated for different viewing distances. MS-SSIM also puts less emphasis on the luminance component compared to the contrast and structure components. In total, MS-SSIM has been shown to increase the correlation between the MS-SSIM index and subjective quality tests. However, the trade-off is that MS-SSIM takes a few times longer to run than the straight SSIM algorithm.
Original paper: https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf
MS-SSIM* is an extension to the original MS-SSIM algorithm proposed by Wang, et al. MS-SSIM* does away with the stabilization constants and defines concrete values for the edge cases. The authors provide data that shows a modest increase in correlation between the MS-SSIM* index and subjective tests, as compared to MS-SSIM. MS-SSIM* is the default algorithm used by the iqa_ms_ssim() method.
Original paper: http://foulard.ece.cornell.edu/publications/dmr_hvei2008_paper.pdf
