Image Quality Assessment Based on SIFT and SSIM - Advances in Image and Graphics Technologies - page 4

Reference
Image

Luminance
Comparison

Luminance

+ Contrast

Contrast
Comparison

SIFT
Feature

Combination

÷

Test
Image

Structure
Comparison

Luminance

SIFT
Feature

+ Contrast

÷

Qual ity

Comparison
Normalization

×

Fig. 2. Diagram of SIFT-SSIM Algorithm

When the two images SIFT feature vector generation, the next step we use the key
feature vector images Euclidean distance as the two key points in determining the
similarity measure. Take a key reference point in the image, and identify it with the
test image Euclidean distance nearest first two key points in these two key points, if
the closest distance divided by the distance is less than the proportion in threshold
value, the acceptance of this pair of matching points. Lower this threshold ratio, SIFT
matching points will reduce the number, but more stable.
In normalization process of comparison, we refer to proportion of SIFT matching
points as one of quality assessment factors. Algorithm performs the following steps:
Algorithm: SIFT-SSIM.

To read a reference image x and a test image y ￿
2￿To

1•To read a

To extract feature points of the two images, numbers of

NumSIFT x ()and NumSIFT y ￿ ()

features are

3￿To To Look for matching feature points￿their numbers is their numbers is

NumMatch x y ￿
4￿To c

(, )

To calculate the percentage to gain features score
(, )/ SIFTScore = NumMatch x y NumSIFT x ()

￿

5￿To c To calculate brightness, contrast, and structure simi-

larity functions of the two images, which are
(, ),(, ),and (, ) lxy cxy sxy , and to get SSIM score

SSIM xy (, ) = lxy cxy sxy (, ) (, ) (, ) ⇅ ⇅

￿

6￿To c To calculate the final quality score

_

SIFT SSIM = SIFTScore SSIM x y ⇅ ( , )