【作者】 赵谊虹；

【导师】 余松煜；

【作者基本信息】 上海交通大学 ， 信息与通信工程， 2007， 博士

【摘要】 图像和视频压缩技术在很大程度上减小了图像和视频的存储容量和传输带宽,促进了图像和视频业务的广泛应用。但压缩也导致图像和视频中出现了编码效应,如块效应、振铃、蚊子效应等。这些编码效应的存在破坏了原有图像的空域平滑性,降低了解码后图像的主观和客观质量。图像和视频的后处理可以显著地消除编码效应,提高图像的主观和客观质量。本文的主要工作是针对消除“DCT+量化”过程所导致的编码效应,提出了一些后处理算法:本文首先研究了DCT域块效应消除算法;然后提出了空域块效应消除算法;接着提出了振铃效应的消除算法;最后提出了与解码器相结合的后处理方案,用于消除视频编码中出现的块效应和振铃效应等编码效应。本文的主要创新成果包括:提出了一种DCT域渐进块效应消除算法(PDCT-BAR)。该算法根据解码DCT系数块的频率特征参数来对偏移DCT系数块中的系数进行限制,以消除块效应。偏移的DCT块中可能存在由块效应引起的高频系数,通过不断构造偏移的DCT块,并且自适应地约束偏移块中的这些高频分量,从而达到了消除块效应的目的。此外,在完成块偏移和系数收缩后,使用量化约束来修剪处理后的DCT系数。与其他使用偏移块的块效应消除算法不同的是,PDCT-BAR算法的每一次偏移过程都构造了具有不同偏移位置的偏移块。实验表明,通过对多个偏移位置块的处理,PDCT-BAR算法不仅可以有效的去除块效应,而且能很好的保存图像本身的纹理细节,处理后图像无论PSNR质量还是视觉质量都令人满意。提出了一种基于DCT域自适应滤波的块效应消除算法(DCT-AFBAR)。在DCT-AFBAR算法中,首先推导出四组用于DCT域滤波的系数矩阵,然后根据解码块的频率特征参数来选取当前DCT块滤波所使用系数矩阵,最后将相应的DCT系数块与选好的系数矩阵进行相乘,即得到消除了块效应的DCT系数块。该算法的优点在于避免了计算复杂度非常高的偏移DCT系数块计算。为了进一步减小计算复杂度,文中还测试了只对部分DCT系数进行处理的情况,DCT-AFBAR算法的性能仍令人满意。实验表明:DCT-AFBAR算法无论是在图像主观、客观质量、还是计算复杂度上都比同类算法有显著的改进。提出了一种基于空域自适应滤波的块效应消除算法(SAFBAR)。SAFBAR通过对DCT-AFBAR算法进行推导和改进,将DCT域滤波转化为空域滤波。SAFBAR算法根据解码块的频率特征参数自适应地选取滤波器参数来进行滤波。为了避免真实边缘被模糊,该块效应消除算法还联合多个块边界的梯度信息来区分块效应与真实边缘。随后文中又对SAFBAR算法进行简化得到MSAFBAR算法。MSAFBAR算法对SAFBAR算法中的滤波器参数进行了简化,并将浮点型滤波运算变为整数型滤波运算。为了避免在DCT域计算解码块的频率特征参数,文中进一步提出使用近似变换或梯度变化信息来计算解码块的频率特征参数。实验表明:在不同比特率下,MSAFBAR算法在PSNR性能均优于其他算法。特别值得注意的是,MSAFBAR算法不但在比特率非常低的时候能提升解码图像的PSNR,而且当比特率较高的时候也能提升解码图像的PSNR质量。结合灰度变换和低通滤波提出了一种新的振铃效应消除算法。该算法首先根据解码块的频率特征参数来选取需要进行振铃效应消除的候选块;然后对块内象素的灰度值进行分析,为该块的处理计算出相应的块级参数,包括一个加权系数和两个目标灰度值;接着通过灰度变换将块内每个象素的灰度值向相应的目标灰度值逼近,并同时采用低通滤波得到平滑结果,最后通过块级加权系数将灰度变换和低通滤波的结果进行加权。为了防止象素被过度处理,还对处理后象素的取值范围进行了限制。实验表明:对于振铃效应明显的图像,经过算法处理后,图像的PNSR质量和视觉效果得到了很大的提升。针对视频压缩码流解码后视频序列中的编码效应,本文提出了与解码器相结合的后处理方案。该方案利用了解码过程中的一些信息,并针对不同帧类型来计算解码块的频率特征参数,然后利用该参数来指导块效应消除和振铃效应消除。与同类算法相比,本文的算法对I、P、B帧类型都取得了令人满意的效果。更多还原

【Abstract】 Image and video compression technology could greatly reduce the image/video storage volume and the transmission bandwidth, thus boost the widely application of image and video. However, the compression technology also introduces coding artifacts, such as the blocking artifacts, the ringing artifacts and the mosquito artifacts etc. These artifacts to some extent destroy the spatial smoothness of the original content, and degrade both the objective and the subjective quality of the reconstructed image/video. In order to alleviate the coding artifacts and improve the quality of the reconstructed image/video, it is essential to combine post-processing with the decoding process of image/video. This dissertation focuses on several algorithms reducing the coding artifacts caused by the conventional“DCT + Quantization”in the image/video compression procedure. Firstly, the deblocking algorithm in the DCT domain; secondly the deblocking algorithm in the spatial domain; then the deringring algorithm and lastly the postprocessing to reduce the coding artifacts in video. The originality innovations of this dissertation includes:Progressive DCT domain blocking artifacts reduction (PDCT-BAR). This algorithm reduces the blocking artifacts by constraining coefficients in the shifted DCT blocks according to frequency characteristic parameters of the decoded DCT blocks. The undesired high frequency components, mainly caused by the blocking artifacts, are constrained by progressively constructing the shifted DCT blocks and adaptively shrinking the high frequency coefficients in the shifted blocks. Then the DCT coefficients are clipped by the quantization constraint set. Comparing with other deblocking algorithms related with shifted blocks, the algorithms in this dissertation constructed shifted blocks with different shifting position in each shifting. Experiments show the PDCT-BAR algorithm could both effectively reduce the blocking artifacts and highly preserve the details in the original image by take the advantage of different shifted blocks. The post-processed images give satisfying PSNR and visual quality.DCT domain adaptive filtering of blockiness artifacts reduction (DCT-AFBAR). First, four groups of filter coefficients matrix for DCT domain deblocking are deduced. Then frequency characteristic parameters are extracted to instruct the selection of the filter coefficients matrix. The selected filter coefficients matrix multiplied with the DCT coefficients and the blocking artifacts is alleviated in the filtered DCT blocks. One advantage of this algorithm is the reduction of heavy computation burden of calculating the shifted blocks. Beside, in order to further reduce the computation complexity, partial process to the DCT block is also tested in this dissertation, and the result is also acceptable. Experiments show: the DCT-AFBAR algorithm outperforms the similar algorithms both in objective/subjective quality and computation complexity.Spatial adaptive filtering of blockiness artifacts reduction (SAFBAR). SAFBAR is the deduction and improvement of DCT-AFBAR, the DCT domain filtering is transformed into spatial filtering. SAFBAR extracts frequency characteristic parameters to instruct the selection of filter coefficients. To avoid blurring of the real edge, this algorithm also combined the gradient information of several block boundaries to distinguish the blockiness and the real edge. MSAFBAR is a modified SAFBAF algorithm, the filter coefficient is simplified. So that the floating point calculation is converted into integer calculation. Furthermore, to avoid the DCT domain calculation of frequency characteristic parameters, approximate transform or gradient variation information are used to approximate the frequency characteristic parameters. Experiments show: MSAFBAR outperforms other algorithms in PSNR under a wide range of bitrates. A ringing artifacts reduction algorithm is proposed, which combines gray level transform and low-pass filtering. This algorithm selects candidate blocks by the frequency characteristic parameters. The parameters for block level, includs one weighting value and two trageted gray level values. The gray level transform approximates the grey level values of pixels to the targeted value; and the low pass filtering gives another smoothed output. Then weighting the output of gray level transform and the low-pass filtering to get the result without ring artifacts. To avoid excessive processing, the result is limited to give the untimate result. Experiments show: the proposed algorithm is especially effective in PNSR and visual quality of the decoded images.This dissertation also propose a scheme of post-processing with decoder to alleviate the coding artifact in the decoded video bitstream. This scheme utilize some informations of decoding process to calculate the frequency characteristic parameters for different frame types, then the parameters are utilized to instruct the deblocking and deringring. The proposed algorithm is satisfying for I, B and P frame types as compared to like algorithms.更多还原