标题: 研究自动回归技术在提升影片播放速率与多重描述视讯编码上之应用
Study on Auto-Regressive Model for Its Applications on Frame-Rate up Conversion and Multiple Description Video Coding
作者: 蔡文锦
Tsai Wen-Jiin
国立交通大学资讯工程学系(所)
公开日期: 2012
摘要: 自动回归分析(auto-regression analysis) 是在统计学上的一种方法,主要是对资料之间的关系建立模型,目的在于了解资料间是否相关,以便观察或预测我们感兴趣的资料。自动回归的技术可用在影像或视讯资料上,对已知的点(pixel)建模去预测不存在的点或重建遗失的点。因此,在一些应用如: 增加影像解析度,增加视讯播放速率(frame rate up-conversion, FRUC),错误隐藏方法(error concealment)等当中,可以看到使用该技术的例子。在现有的方法中,有的只利用到资料在空间上的关系、有的则只利用到时间上的关系、有的虽然同时利用空间与时间的关系,但所建出的回归模型太复杂,需要求解的回归系数太多,也往往需iteration多次才能得到较佳解,不适用在实际的应用中。有鉴于此,本计画拟进行的研究之一,是改善目前在FRUC应用中的回归技术。
另一方面,多重描述编码法(Multiple Descriptor Coding,简称MDC)是一种将影片切割成多个子影片(或称描述档,description)的视讯编码法,非常适合用于网路上的影音串流的应用中。由于它将高频宽的影音资料切割成多个低频宽的描述档,使得低阶设备也能参与上载或下载较低频宽的描述档,而高阶设备则能藉着接收多个描述档来解出较高品质的影片。此外,由于多重描述编码所切割出来的描述档是可以被个别独立解码的,且各描述档在传输时可走不同路径,因此,有封包遗失时可藉由其他描述档的内容来加以修复,大幅减低封包遗失对影片品质的影响。此编码法虽然能为设备的异质性与网路的封包遗失提供解决之道,然而编码效能差却是它主要的问题所在。在多重描述编码中,因为资料间的关系被切割分散在不同描述档中,无法充分利用,造成整体编码的资料量远高于一般标准编码所产生的资料量。因此,本计画拟进行的另一部份研究,是将回归分析技术用于多重描述编码中,以达到减少多余的资料量,并维持容错能力的目的。
此研究计画中,我们预计以两年为期完成以下两个目标:1.改善目前在增加视讯播放速率(frame rate up conversion)应用中的回归技术; 2.将回归技术用在多重视讯编码上。第一年的研究着重在减少使用回归技术之计算复杂度,并进一步提升其所产生的画面品质,使其更适用在实际系统中。我们拟为资料的空间与时间相关性分别建构不同的回归模型,再设计一个选择适当回归模型的机制,来达到减少计算复杂度并维持原有画面品质的目的;同时我们也将利用动态改变回归模型中使用的training window大小,来进一步提升画面品质。第二年的研究着重在将回归技术用在多重描述视讯编码上,希望利用少量的 redundancy,达到较佳的资料容错能力。由于回归技术能为资料间的关系建模,我们的作法是在描述挡内加入回归系数,以保留资料间的相关性,由于回归系数的资料量远小于原始资料的量,因此我们预期所使用的方法能用少量的 redundancy,达到不错的容错能力。
Auto-regressive (AR) analysis is a technique to model the relationship between data, and therefore, it can be used to interpolate and recover “dirt” areas in image sequences. AR model has been applied in many applications, such as frame-resolution up scaling, frame-rate up conversion, error concealment, super-resolution, video data forecasting, etc. Current methods utilizing AR model take either spatial or temporal relationship into considerations only. Even though some methods take both relationships into account, they suffer from large computation complexity and thus are not practical for real applications. Therefore, one of our project goals is to improve the AR model used in frame-rate up conversion.
Multiple description coding (MDC) is a technique of striping a video sequence into two or more descriptions in such a way that each description is independently decodable. MDC technique is very suitable in streaming application for handling heterogeneous devices and dynamic network conditions. By means of striping a video sequence of high bit-rate into more descriptions with low bit-rates, devices can choose the number of descriptions to receive according to their capability; besides, if some descriptions are lost, the playback can continue with gradual quality reduction instead of complete failure. Although MDC provides a good solution or streaming application, it suffers from the problem of poor coding efficiency. Hence, the other goal of this project is to propose a MDC method which incorporates AR model to utilize little redundancy and provide error robustness for coded video sequences.
There are two goals that we would like to achieve in two years: 1. Study on improving AR model used in frame rate up-conversion, 2. Study on designing a MDC approach based on AR model. In the first year, we plan to separate temporal and spatial relationships in AR models and adopt an adaptive selection mechanism to reduce the computational complexity. We also plan to dynamically select appropriate training window sizes used in the AR model in order to further improve the produced frame quality. In the second year, we will focus on designing a MDC method which incorporates the parameters of AR model in to produced descriptors. By maintaining data relationship using AR parameters in MDC, we expect that this method can use little redundancy to greatly improve the error robustness.
官方说明文件#: NSC100-2628-E009-026-MY2
URI: http://hdl.handle.net/11536/98265
https://www.grb.gov.tw/search/planDetail?id=2391097&docId=380240
显示于类别:Research Plans