1、PDF外文：http:/ B 外文参考文献 Lossless Image Compression with Lossy Image Using Adaptive Prediction and Arithmetic Coding Seishi Taka and Mikio Takagi Institute of Industrial Science,University of Tokyo Abstract &
2、nbsp; Lossless gray scale image compression is necessary in many purposes, such as medical image, image database and so on. Lossy image is important as well, because of its high compression ratio. In this paper, we propose a Lossless image compression Scheme using a lossy image generate
3、d with PEG-DCT scheme. Our concept is, send a PEG-compressed lossy image primary, then send residual information and reconstruct the original image using both the lossy image and residual information. 3-dimensional adaptive prediction and an adaptive arithmetic coding are used, which fully uses the
4、statistical parameter of distribution of symbol source. The optimal number of neighbor pixels and lossy pixels used for prediction is discussed. The compression ratio is better than previous work and quite close to the originally Lossless algorithm. Introduction Tod
5、ay there are many studies on image compression, particularly on lossy and very low bit rate compression. For image database, such high compression ratio is Important for storage and also for quick transmission,but to deal with various kinds of users demand, Lossless image transmission is
6、 indispensable. In this paper, we propose an effective Lossless compression algorithm for gray image using lossy compressed image. The lossy compression scheme uses the Joint Photographic Experts Group discrete cosine transform (PEG-DCT) algorithm as the lossy coding algor
7、ithm. First we search the similar pairs of pixels (conlexts), according to their neighbor pixels. For such pixels which have contexts,we predict their values from the contexts and the neighbors. On the other hand, for each pixel which doesn't have its context pairs, we calcu
8、late the edge level according to the difference of adjacent pixel values. For each edge level of pixels, we calculate the predictive coefficients of linear combination under the least square error criterion. Not only the pixels which have already processed but also the pixels of the lossy image is u
9、sed for prediction. For every pixel, the difference between predicted value and real value is calculated, and the difference is converted to anon-negative value before being encoded, according to their distribution. In entropy coding stage,we use the arithmetic coding. It is mad
10、e adaptive,and initial error distribution is given only by one para meter, which is specific for each edge level's statistical distribution. The pixels belonging to the difference edge levels are encoded independently. Experimental results and good performance are shown
11、. Like the other LPL(L0ssy Plus Lossless) approaches, our compression ratio is less than that of originally Lossless scheme,but the difference is slight. Of all things, however, users get the great merit That they can browse the image before the Lossless decompression.Many such schemes have been pro
12、posed in the literature, but most of them treat the lossy image and its Lossless residual as independent symbol source. One of the exceptions is Mem on s algorithm6. We utilize the lossy data thoroughly,and much better result is obtained. 1.1 Pixel estimation
13、 Normally the image data is scanned along the scan-line direction.In figure 1. current pixel is processed pixels are . Ordinal pixel estimation method predicts the NFL,value of current pixel using PI . .P 4 . Then calculate the prediction error e= - . Normally the li
14、near combination is used for prediction as follows,where TI . . .T4 are the coefficients. This Figure 1: Current pixel is the extrapolative prediction.processed neighbor pixels = ctptptptp 44332211 (1) Us
15、ually, the zero-order entropy of set e is lower than that of set z.Therefore, after entropy coding scheme such as Huffman codingl or Arithmetic coding2 or Lempel-Ziv coding3, data size is reduced. Principally, we also use the linear combination like equation (1) for prediction, &
16、nbsp;but the process is more adaptive than normal prediction method. And we use more neighbor pixels (up to ten), also using the pixels of lossy image and prediction error e is converted to another form before encoding. 2.1 Grouping the pixels Each image
17、pixel has different property under certain criterion. From a point of image compression, grouping similar pixels and encode them together causes x xxeffective result. For grouping the pixels, we use the Q value: Q= | pppp|pp|pp|15141312 &nb
18、sp; (2) Using this Q value, we classify each pixel into several groups according to table 1 Table 1:Grouping table
19、Figure 2: (a)Original image Girl and (b)JPEG compressed image(qua1ity value=5) Figure 3: (a)Image of Q value, (b)Image of prediction error of simple prediction Figure 3(a)shows the Q value and (b) shows the error of simple prediction.
20、As can be seen from them, the value Q correlate closely with the prediction error. Therefore the prediction coefficients are calculated independently in each group. 2.2 Context search Table 2 shows each group s final zero-order entropy of prediction result of image Girl . Obviously,the upper groups are more difficult to be compressed than the lower groups. We use the context-based prediction method to deal with such upper
