面向视频压缩的显著性协同检测JND模型

Improved Just Noticeable Distortion Model Based on Saliency Detection for Video Coding

为了更好的将人眼感知特性用于视频压缩系统,提出了一种改进的基于显著性协同检测的恰可察觉失真模型(Just Noticeable Distortion,JND).该模型通过像素域和变换域下联合建模计算得到的最优JND模型,基于上下文感知的显著性算法得到相应的显著图,并将检测结果用于JND模型权值分配.提出的JND残差滤波器可以嵌入到HEVC视频编码框架中.实验结果表明:在全I帧配置下,提出的算法编码结果与HM16相比,在视觉主观感知质量一致的情况下,平均码率可节省10.7%.

A improved just noticeable distortion（JND） model based on saliency map detection algorithm is proposed for video coding in order to apply feature of visual perception to the video compression system. The optimal JND model is calculated by setting modle of pixel-based JND with the hybrid. The saliency map concluded by context aware saliency detection modle is used for the weight distribution of JND model. The proposed model and residual filter can be integrated into the framework of HEVC, which is useful for quantifying video data. The experimental results shows that in the case of visual subjective perceptual quality, the average bitrate reduction is 10.71%, compared with HM 16（all Intra profile）.