应用log-Gabor韦伯特征的图像质量评价

考虑人眼对亮度的感知符合韦伯定律的特点,本文利用log-Gabor滤波器模拟人眼对图像的感知过程,提出了一种新的log-Gabor韦伯特征,以便保留不同尺度的符合人眼感知的结构信息。基于此,还提出了一种应用log-Gabor韦伯特征的图像质量评价方法。首先将待评价失真图像和参考图像从RGB空间转换到YIQ颜色空间,分离亮度分量和颜色分量。然后利用log-Gabor韦伯特征和梯度特征计算亮度分量失真,并结合颜色分量的失真,得到失真图像与参考图像的局部相似度图。最后利用修正的CSF函数,对局部相似度图进行加权,得到图像质量评价指标。在LIVE、CSIQ和IVC3个图像库上的实验结果表明,本文方法与人眼主观感知有很好的一致性,而且相对于其他方法,表现更加稳定。本文方法在3个图像库上的加权Spearman秩相关系数(SROCC)为0.949 8,Kendall秩相关系数(KROCC)为0.802 6,Pearson线性相关系数(PLCC)为0.943 8,相比对比方法有显著的提高。

Image quality assessment using log-Gabor Weber feature

As human eye perception for the brightness accords with the Weber's law, this paper uses the log Gabor filter to simulate the human eye perception for an image and proposes a new log Gabor Weber characteristics to keep the structural information interested by human for different scales. To assess the image quality more effectively, a new image quality assessment method was proposed by using log-Gabor Weber feature. The log-Gabor filter and Weber's law were used to obtain a new feature named log-Gabor Weber feature(LGW). Firstly, the distorted image and reference image were transformed from the RGB color space into a YIQ color space to separate the luminance component and the chromatic component. Then, the LGW feature and gradient feature were used to calculate the distortion of luminance component. Furthermore, the distortion of chromatic component was integrated to get the local similarity map between distorted image and reference image. Finally, a modified CSF pooling strategy was applied to the overall local similarity map to obtain the final image quality index. The experimental results on three benchmark image databases, LIVE, CSIQ and IVC, indicate that the proposed method owns a good consistency with human subjective perception and it has a more stable performance as compared with other state-of-the-art methods. The weighted Spearman Rank Order Correlation Coefficient(SROCC), Kendallrank-order Correlation Coefficient(KROCC) and the Pearsonlinear Correlation Coefficient, PLCC) values on three databases by the proposed method are 0.9498, 0.8026 and 0.9438,respectively, which notably outperform other methods.