低码率下基于ROI与JND的分级量化方法  

New Method for Graded Quantization Based on ROI and JND at Low-bitrates Coding

作　　者：喻莉[1] 冯慧[1] 张军涛[1] 左雯 王宁 

机构地区：[1]华中科技大学电子信息与工程系,湖北武汉430074 [2]中兴通信股份有限公司,广东深圳518057

出　　处：《电视技术》2013年第17期33-35,共3页Video Engineering

基　　金：国家实验室基金项目(P080010);中芬国际合作项目(2010DFB10570)

摘　　要：传统基于人眼感兴趣区域(ROI)的分级量化模型,将视频帧划分为ROI区域和非ROI区域,对相应区域用不同的量化参数(QP)进行量化,以提升视频的主观质量。而该模型没有考虑ROI区域的内部特性,不能很好地符合人眼视觉特性(HVS)。针对低码率条件下以人脸为主体的桌面视频、手持终端等场景,提出一种基于ROI和恰可观测失真(JND)的分级量化方法。JND模型表明边界区域相对平滑区域能够隐藏更多的失真,利用该属性检测出ROI区域(即人脸)中人眼更感兴趣的边界部分(例如眼睛、鼻子、嘴巴等),据此建立基于ROI与JND的分级量化模型,指导各区域的量化。实验结果表明,针对低码率视频的应用,与传统分级量化方法相比,本文所提方法在相同码率条件下能明显提升视频的主观质量。According to the traditional graded quantization model, the frame can be divided into ROI and non - ROI in terms of the region of the eye interest in order to improve the subjective quality of the video with different quantitative parameters（ QP）for the correspondingly specific region. However, the model can not reflect Human Visual System（HVS） well without the consideration of the internal character of ROI. Therefore, this paper presents a method for graded quantization based on ROI and Just-Noticeable-Distortion（JND） ,which is mainly used for video calling at low-bit rates coding in the settings such as Desktop Video,Handheld terminal with the subject of human face. JND model shows that the boundary area can conceal more distortion than the smooth region, thus according to this characteristic, the boundary in the ROI region （human face）can be detected for instance, eye,nose, mouth, and so on. And the graded quantization model is proposed with the basis of ROI and JND,which can direct the quantization of different regions. Compared with the traditional graded quantization modal,the results of the experiment demonstrated that for the application of the low bit video,the method could improve the subjective video quality at the same bit rate.

关 键 词：人眼感兴趣区域 恰可失真模型 分级量化 低码率 

分 类 号：TN919.8[电子电信—通信与信息系统]