| 摘要: |
| 目的 针对高斯模糊这种失真类型，提出一种盲评价算法。 方法 从权威的图库中选取高斯模糊图片，提取梯度，并对梯度图进行快速傅里叶转换（FFT） ，得到频谱，对原图、梯度图、频谱进行运算，提取出边缘强度、方差、梯度熵，作为每幅图的特性向量。通过 GRNN 构建可以计算出图片的差分平均意见得分（DMOS）值，即输入特征值，输出计算 DMOS 值。 结果 该算法的Spearaman 秩相关系数（SROCC）达到了 0.9086， Perason 线性相关系数（PLCC）高达 0.9033；与一些常见的算法相比，所运算产生的 SROCC 和 PLCC 值也更高。 结论 使用 CSIQ 与 LIVE 图库的高斯模糊部分，在 Matlab 的环境下进行运算后得到的结果表明，计算产生的 DMOS 与由人评判产生的 DMOS 值相似度高，与眼睛判断结果较为接近。 |
| 关键词: 模糊图像质量 盲评价 梯度 GRNN |
| DOI： |
| 分类号:TS801.3； TS807 |
| 基金项目: |
|
| Blurred Image Blind Assessment Based on GRNN |
|
YUAN Pu-kang1, KUANG Sheng-kun1, WANG Qiang2, TIAN Quan-hui3
|
|
1.University of Shanghai for Science and Technology, Shanghai 200093, China;2.Hangzhou Dianzi University 310018,China;3.Shanghai Publishing and Printing College, Shanghai 200093, China
|
| Abstract: |
| A new blind assessment algorithm was put forward for Gaussian Blur. This method selected Gaussian Blur images from an authorized image library, extracted the gradient and conducted Fast Fourier Transform (FFT) and then got a frequency spectrum. It then calculated the initial image, the gradient image and the frequency spectrum, and extracted the edge intensity, the variance and the information entropy as characteristic vectors of each image. GRNN (characteristic vector) was used to build the blur image quality evaluation model which could input characteristic vector and output the calculated Different Mean Opinion Scores (DMOS). Compared with other common algorithms, this method had relatively higher Spearman’s rank correlation coefficient (SROCC) at 0.9086 and Pearson linear correlation coefficient (PLCC) at 0.9033. The results of the calculation, which utilizes the blur image, set of CSIQ and LIVE database in the Matlab environment indicates the calculated DMOS with the new algorithm has high similarity with the DMOS of subjective judgment and is nearly close to human visual judgment. |
| Key words: image of blurred image blind assessment gradient general regression neural network |
