面向HEVC的恰可察觉编码失真模型

Just noticeable coding distortion model for HEVC video coding

为进一步提高现有视频编码技术的压缩效率及解码重建图像的主观视觉感知质量,在现有人眼恰可察觉失真(JND,just noticeable distortion)模型的基础上,提出了恰可察觉编码失真(JNCD,just noticeable coding distortion)模型。首先,通过主观实验,对恰可察觉梯度幅值差异(JNGD,just noticeable gradient difference)进行了研究,分析其变化规律并建立JNGD模型。使用全变分(TV,total variation)方法将图像分解为结构图和纹理图后,分别求取其梯度信息得到结构梯度图和纹理梯度图,利用JNGD模型分别滤除结构梯度图和纹理梯度图中的人眼不可察觉的梯度幅值;其后,分析了人眼感知对于不同梯度幅值的编码失真敏感性,设计了梯度幅值与JNCD值的主观实验,得到两者的关系模型;最后,考虑人眼对图像中的边缘、平坦和纹理3类区域失真感知程度的差异性,利用滤波后的结构梯度和纹理梯度信息将图像划分为上述3类区域,最终建立整幅图像的JNCD模型。为验证本文提出的JNCD模型的可靠性,在高效视频编码(HEVC)标准测试平台上进行的模型验证结果表明,在本模型指导下的编码其解码重建图像获得了较好的主观视觉效果,可为人眼视觉感知冗余的分析及感知编码的改进提供依据。

To improve the efficiency and perceptual quality in video coding, this paper proposes a model of just noticeable coding distortion （JNCD）, which considering the human visual perception redundancy and the unreasonable factors of the existing just noticeable distortion （JND） models in coding process. First, we design a psychophysieal experiment to analyze the just noticeable gradient difference （JNGD）, and build JNGD model to filter the gradient components which are imperceptible by human eyes. We use total variation （TV） to decompose an image into structural image and textural image,and calculating their gradient respectively, using JNGD to filter out the imperceptible gradient components in each gradient image. Second,the human eye sensitivity of different gradient magnitudes is analyzed to model the rela- tionship between the human eye perceptible gradient magnitude and JNCD. Finally, considering the perceive difference of human eye in edge,flat and texture regions of an image,we adjust the JNCD value in each region and establish JNCD model of the whole image. To verify the efficiency of the proposed JNCD model, we test the JNCD model on the high efficiency video coding （HEVC） platform. Compared with comparison schemes, the proposed model shows its advantages in subjective visual effect, which means that it is helpful for analysis of human visual perception redundancy and the relevant perceptual video coding.