基于自然图像统计的无参考图像质量评价

No-reference image quality assessment based on statistical model of natural image

为了解决无参考图像情况下的质量评价问题,基于变换域的自然图像统计模型,提出一种适用于白噪声、高斯模糊、JPEG2000压缩等失真类型的无参考图像质量评价方法.利用自然图像Contourlet变换域子带均值间的线性关系,用失真条件下保持不变的低频子带均值预测未失真的高频子带均值,以高频子带预测均值与实际均值之间的差异计量图像的失真.结合人类视觉系统特性,对不同尺度、不同方向子带及子带内不同区域进行选取与加权,综合得到对失真图像的客观质量评价.实验表明,该质量评价测度与主观质量评价有较好的一致性,且在性能上优于峰值信噪比(PSNR).

A new no-reference image quality assessment method was proposed based on the statistical model of natural images in order to assess the distortion of white noise,Gaussian blur and JPEG2000 compression.For natural images,the means of Contourlet subband coefficient amplitudes （MSC） decreased approximately linearly with scale index.Then MSC of high frequency was predicted from MSC of low frequency,which was not badly affected by distortion.The distortion was evaluated using the difference between the predicted means and the real means in distorted image.The metric of subband distortion for different scales,different directions,and different regions was weighted and pooled into a quality score considering characteristic of human visual system.Experimental results showed that the method was consistent with subjective assessment,and exceeded the performance of peak signal to noise rate （PSNR）.