Graphics Reference
In-Depth Information
1.2
Full-Reference Methods Based on Structure Similarity
Because of applying HVS model to image quality assessment exist some problems,
researchers have proposed image quality assessment model based on structural simi-
larity. They consider that natural images have a specific structure, the pixels have a
strong affiliation, and the affiliation reflects the structural information of visual scene.
Zhou Wang and A.C. Bovik proposed image quality assessment method based on
structural distortion, referred to SSIM [3]. In the method the comprehensive quality
assessment model was proposed by using mean, standard deviation and unit standard
deviation to represent brightness, contrast and structural similarity, it is the important
landmark of IQA. The method deems that illumination is independent of object struc-
ture, and the light is mainly from changes in brightness and contrast. So it separates
brightness and contrast from image structure information, and combines with struc-
tural information to image quality assessment. The method actually steers clear com-
plexity of natural image content and multi-channel relations to evaluate structural
similarity. It advantage to its lower complexity and wider applications. However, the
algorithm only considers the structure of an image except to image features, but fea-
tures of each image patches have different effects on image quality.
Subsequently, Zhou Wang and others also proposed a multi-scale SSIM [4], the al-
gorithm obtained better results than a single scale. And the method introduced
weights of information content to SSIM was proposed. In the algorithm, weights were
calculated the proportion of information content of patches to the whole image in both
reference and distorted image.
Lin Zhang et al proposed FSIM [5], two features of phase coherence and gradient
were applied to calculate local similarity mapping. In pooling strategy of quality as-
sessment, phase coherence was used again as a weighting function, because it can
reflect local image in the perception of the importance of HVS well.
Lin Zhang et al also proposed RFSIM [6], in this method first-order and second-
order Riesz transform were used to characterize the local structure of image, while
Canny operator of edge detection was used to produces pooling mask of quality score.
Therefore, we propose combination strategy of feature matching and structural si-
milarity, experiments approved that it can improve SSIM algorithm and even be com-
parable to other state-of-art full reference method.
2
Design of Full-Reference Algorithm Based on SIFT-SSIM
2.1
SIFT and SSIM
SIFT (Scale-invariant feature transform) algorithm was proposed by D.G.Lowe in
1999 [7], and in 2004 it was improved [8]. Later Y.Ke improved its partial descriptors
by using PCA instead of histogram. SIFT algorithm is a local feature extraction algo-
rithm, and it wants to find extreme points in scale space, extract location, scale and
rotation invariant. SIFT features are local features of image, and remain invariant of
rotation, scale, brightness, also maintain a certain stability of angle, affine transforma-
tion, and noise.
Search WWH ::




Custom Search