Abstract
對於消於區塊效應(Blocking Artifacts)而言，後處理是非常重要的。所有的以區塊為基礎的壓縮方式，都會在區塊邊界處產生區塊效應，特別是低位元率的時候。而區塊效應似乎是轉換壓縮的主要缺點，也是低位元率壓縮(Low Bit-Rate Coding)的主要限制。以區塊為基底的轉換壓縮方式，已經廣泛的被使用在各式各樣的應用上，包括影像處理及視訊訊號。然而，常常會在解壓縮回來的影向上產生區塊效應。有兩種常用來去除區塊效應的方式。第一種是試著保留更多關於區塊相互之間的影像資訊。這種方式可以藉由增加在轉換係數上區塊之間關係或送額外的訊息給接收端藉使用更多的位元率。一些這樣的演算法可以展現不錯的結果，雜訊比(PSNR)可以增加一至二個分貝(dB)。然而，這些方法，並不相容於現有的產業用標準，因為他們修改了現有標準的根本轉換方式。而另一種方式則是使用後處理的技巧，最典型的應用就是區塊邊界濾過器(Block-BondaryFilter)。他們因為可以相容於現有的標準而且不需要送額外的訊息，所以受到大家的重視。然而這些方法，致力於使影像的視覺品質提高而非重建影像。所以失真度(Distortion)，雜訊比(PSNR)可能在處理過之後反而變的更差。而至今還沒有一種廣為接受的衡量標準來測度區塊效應。所以，影像品質僅能以人眼來判定。在這篇論文中，我們提出一個後處理的新方法。我們使用最小平方法(Least-Squares Estimator)來消除區塊效應。因為以區塊為基礎的轉換壓縮方式忽略了相鄰區塊之間的關連性(Correlation of Adjacent Block)。這樣很容易在區塊邊界處產生假的邊界(False Edge)。另外，我們也考慮到直接使用濾過器(Filter)會造成影像模糊的現象(RingingEffect)。我們的演算法只修改邊界處，而非改變整個區塊的像素值(Pixel Value)。對於每一個區塊，我們分別對水平方向及鉛直方向做平滑處理。這樣其實已經有點表面近似(SurfaceFitting)的味道。我們先分別找到一些取樣點，當然我們先避開有含邊界(Edge)的點，再把這一些取樣點套上最小平方法得到一條曲線。之後，我們就可以對應每一個點得到對應的像素值。我就可以把邊界的值換成對應曲線上的值。這個方法易於執行。而處理過後的影像也的確看起來比較好。他的雜訊比(PSNR)也略有提升約 0.2 分貝左右。Postprocessing is very important for eliminating the blockeffect of acoded image. For all of the block-based codingalgorithmsit creates some artifacts across block boundary.Theblocking-artifacts seem to be one of the majordisadvantage oftransform coding and is also the main limitation forlow bit ratecoding. Block-based transform coding is oneof the most popularmethods and is widely used in many applications,including stillimage and video. However, it often produces blockartifacts inthe reconstructed image.Two common strategies are used to reducethe blocking artifacts. Onetries to retain more information ofthe interblock correlation. Thiscan be done by adding theinterblock correlation in the transformationcoefficients orsending extra information to the reciever withmore bit rates.Some of these algorithms show great improvementson increasingPSNR about 1~2 db. However, these methods are notcompatible withthe current industrial standard because they modifythe basictransformation of the existing standard. The other strategyis toutilize the postprocessing technology. Typical examplesareblock-boundary filter. Theybecome noticeable because theyare fully compatible with the existingstardard, and only needthe decompressed image without additionalinformation. However,these methods aim at the image visual qualityinstead of thereconstruction of the image. So the distortion maybe enlargedafter that process. There is still not a well-definedcriterionto measure the degree of blocking artifacts. So, theimagequality is judged by human eyes.In this thesis, we proposea new method on postprocessing. We use a least square method toreduce blocking. Because these block-based transform codingmethods ignore the correlation of theadjancent blocks. It iseasy to form the false edge between the blockboundary. We alsoconsider that filtering algoritms may cause the ringingeffectsto blur the coded image. So our alorithms try to modify onlytheblock boundary rather than the whole block. For each block,we selectsome sample points in this block. Then we apply theleast-squares estimatorto get an approximate curve, and wesubstitute the pixel value of theboundary point with thecorresponding value of the curve. This method iseasy toimplement and we found that the postprocessed imageobviouslylooks less blocking artifacts in our simulationresults.Simulation Results show that we can improve the PSNRabout0.2 db. Also, the postprocessed images looks better thanthe original.