table of Contents
Motivation : that of having a certain similarity compression produces a dummy structure (pseudo structures) block, and a different degree of compression based on the generated dummy structure. Then, we can detect the pseudo structural similarity, to assess the quality of the compressed image.
Detection Method : The maximum degree of compression of the compressed image, to obtain most distorted image (MDI); then calculate similarity before and after compression, i.e., pseudo structural similarity (PSS). If the compressed image itself is of poor quality, then the similarity will be high.
Significance : This method is very effective not only for the natural image compression, but also very effective screen content image (SCI).
This paper considers JPEG compressed images.
[This article appears to be mainly consider blocking effect, because the authors emphasize dummy structure appears in the block edges. In fact, there can be considered within the block blur]
1. Technical details
1.1 to give MDI
First of all, the significance of the PSS is to use mass coordinate the other direction, enabling the blind IQA:
Specifically, the MDI is through the MATLAB imwritefunction, set to 0 mass obtained by compressing.
1.2 determination dummy structure, the dummy structural similarity calculated
[14] points out: a natural image angle (corner) distribution is irregular. But for JPEG compressed images, the angle becomes regular. This is because the introduction of a large compression false corners, and are concentrated in the edge of the block. [14] angle is accounted for by the law, to characterize the degree of compression distortion.
Determining simple manner herein but grossly inaccurate []: As long as the detected angular distribution \ (8 \ times 8 \) on an edge, then it is determined that the dummy structure; otherwise determined to be normal structure. In this way, we can get a pseudo-structure:
Can be seen from the figure, the more intense compression, when the quality is poor, and the structure of the image compression artifacts MDI overlap the more (red for coincident points).
The method of detecting the angle by [17].
Further, PSS is the number of coincident dummy structure divided by the number of the dummy structure of MDI.
2. Experimental
Experimental results are not optimal, but SOTA and methods [7] comparable. Note that on the same test SCI good.
Furthermore, the authors will also be used to detect the general idea of this distortion. Approach is: In some NR feature-based method, based on the PSS as a new feature. As can be seen from Table 2, PSS usually characterized NR significantly improve the performance of the feature-based methods.