通信專業(yè)文獻(xiàn)及翻譯--基于jpeg2000技術(shù)，融合dsp和fpga的可糾錯圖像傳輸系統(tǒng)（英文）

1、Mixed DSP/FPGA implementation of an error-resilient image transmission system based on JPEG2000 Marco Grangetto, Enrico Magli, Maurizio Martina, Fabrizio Vacca CERCOM - Center for Wireless Multimedia Communications Dip

2、artimento di Elettronica - Politecnico di Torino Corso Duca degli Abruzzi 24 - 10129 Torino - Italy grangetto(magli)@polito.it martina(vacca)@vlsilabOl.polito.it Ph.: +39-011-5644195 - FAX: +39-011-564099 Abstract This

3、 paper describes a demonstrator of an error- resilient image communication system over wireless packet networks, based on the novel JPEG2000 stan- dard. In particular, the decoder implementation is ad- dressed, which

4、 is the most critical task in terms of complexity and power consumption, in view of use on a wireless portable terminal for cellular applica- tions. The system implementation is based on a mixed DSP/FPGA architecture

5、, which allows to parallelize some computational tasks, thus leading to eficient sys- tem operation. 1 Introduction ’ Nowadays, there is a growing interest in the end- to-end transmission of images, especially motivat

6、ed by the short-term deployment of next generation mobile communication services (UMTS-IMT2000). However, transmission in a networked, tetherless environment provides both opportunities and challenges. The wire- les

7、s context implies that the data may undergo bit er- rors and packet losses, making it necessary to foresee error recovery modalities. It is thereby necessary that image communication techniques are provided with the

8、ability to recover, or at least conceal, the effect of such losses. The forthcoming JPEG2000 image com- pression standard has been designed to match these requirements, and embeds some error detection and concealment

9、 tools. This paper addresses the development of a demon- strator of an error-resilient JPEG2000 1 1 1decoder im- plementation for image communication over a lossy packet network. The robustness to packet erasures is

10、 achieved by combining the flexibility of the JPEG2000 framework with the powerfulness of source-channel 0-7803-7 147-X/01/$10.000200 1 IEEE adaptive, optimized Reed-Solomon codes. The de- coder implementation is pa

11、rticularly significant in the context of wireless portable terminals for next- generation cellular systems, where the limited power budget and available dimensions impose severe con- straints on the design of a multi

12、media processing sys- tem. 2 System overview the functional units of the implemented system. 2 . 1JPEG2000 image compression JPEG2000 is the novel IS0 standard for still im- age coding, and is intended to provide inn

13、ovative solu- tions according to the new trends in multimedia tech- nologies. At the time of this writing, the standard is in advanced publication stage; the Final Commit- tee Draft [l] is the most recent JPEG2000 d

14、escription publicly available, which our implementation conforms to. JPEG2000 not only yields superior performance with respect to existing standards in terms of com- pression capability and subjective quality, but a

15、lso nu- merous additional functionalities, such as Iossless and lossy compression, progressive transmission, and error resilience. The architecture of the JPEG2000 is based on the transform coding approach. An image

16、may be divided into several sub-images (tiles), to reduce memory and computing requirements. A biorthogo- nal discrete wavelet transform (DWT) is first applied to each tile, whose output is a series of versions of th

17、e tile at different resolution levels (subbands); then, the transform coefficients are quantized, independently for each subband, with an embedded dead-zone quantizer. Each subband of the wavelet decomposition is di

18、vided into rectangular blocks (code-blocks), which are in- In the following we provide a brief description of 1330 code the received bitstream makes the use of a Reed Solomon FPGA implementation very attractive. The J

19、PEG2000 decoder module, entirely implemented on DSP, is composed by four main blocks: syntax parser, entropy decoder (EBCOT), uniform scalar dequan- tizer, and inverse wavelet transform. Moreover, two additional task

20、s, devoted to communication manage- ment between DSP, FPGA and a personal computer, have been introduced. 3.1 Syntax parser The parser is the functional block that interfaces the JPEG2000 decoder with the RS decoder.

21、 It re- trieves RS decoded packets, and extracts from the compressed JPEG2000 bitstream all the relevant in- formation needed to perform image reconstruction. Firstly, the bitstream main header is read, which con- t

22、ains information on the parameters used during the encoding process (e.g. image size, wavelet filter used, number of decomposition levels, quantization thresh- olds, and so on). After that, tile headers are read, whi

23、ch provide information specific to each image tile. Finally, each packet contained in the bitstream is read, and the data and parameters of each codeblock are ex- tracted, and fed as inputs to the EBCOT decoder. 3 .

24、 2EBCOT Right after the bitstream syntax parser, the sub- sequent stage in the JPEG2000 decompression chain is the entropy decoder (EBCOT). From an algorith- mic point of view, EBCOT is a block-based bitplane encoder

25、 followed by a reduced complexity arithmetic coder (MQ). It subdivides each wavelet subband into a disjoint set of rectangular blocks, called code-blocks. Then the compression algorithm is independently ap- plied to

26、every code-block. The samples of every code- block are arranged into so-called bitplanes. To decode a code-block, EBCOT always starts from the most sig- nificant bitplanes, and then moves towards the least significan

27、t ones. The compressed information of every code-block is then arranged in several quality layers, to create a scalable compressed bit-stream. Concep- tually, each quality layer monotonically increases the knowledge

28、of samples magnitudes, i.e. increases the quality of the reconstructed image. Formally, EBCOT is made of three main steps, namely Significance Propagation (SP), Magnitude Re- finement (MR), and Clean Up (CL). Each o

29、f the above steps can resort to four decoding primitives, namely Zero Coding, Sign Coding, Magnitude Refinement Coding, and Run Length Coding. The bitplane vis- iting order follows the sequence SP - MR - CL: it is w

30、orth noticing that every sample of a given code-block is processed in just one of the three steps. As far as computational complexity is concerned, CL demands the largest effort during the decoding of the most sig- ni

31、ficant bitplanes. As SP steps are applied, an in- creasing number of samples become significant, and are inserted in a list of MR-ready samples. Progres- sively, the load required by MR steps grows, making the decode

32、r efficiency directly dependent on the MR and CL optimization level. During the development of the EBCOT decoder block, particular care has been posed on the design of agile data structures, particu- larly suited to

33、DSP optimized C code of MR and CL steps. 3.3 Uniform scalar dequantizer According to [l], the quantization method sup- ported by JPEG2000 is called scalar uniform. Uniform scalar dequantization can be simply accompl

34、ished by means of a single multiplication for each wavelet coef- ficient. 3.4 Inverse wavelet transform The discrete wavelet transform can be evaluated by means of a convolution-based kernel, or a lifting-based ker

35、nel, this latter being the default transform kernel employed in JPEG2000. It has been demonstrated [4] that the lifting scheme may run up to twice as fast as convolution. The wavelet transform has to be per- formed o

36、n both image rows and columns, in order to obtain a separable two-dimensional subband decom- position: JPEG2000 performs first the columnwise, and then the rowwise filtering. The default filter used for lossy compres

37、sion is the well-known DB(9,7): since it does not have rational coefficients, particular care ought to be posed to the effects of finite precision rep- resentation [5]. Due to the use of a fixed point TI TMS320C6201

38、DSP, a detailed study of internal data representation has been performed. Experimental re- sults shows that excellent perceptive quality can be achieved recurring to 9 fractional bits for filter coeffi- cients. In or

39、der to optimize the dynamic range around zero, a DC-shift is foreseen by the standard, as the DC component could lead to an excessive growth of the dynamic range of low-pass subband coefficients. Moreover, the low-pa

40、ss filters can keep the samples in a fixed range, provided that a unitary DC filter gain is guaranteed. The joint effect of DC component sup- pression and unitary gain ensures that range require- ments are fulfilled

41、during the whole wavelet transform. 3.5 Adaptive Reed-Solomon packet de- The de-interleaving RS decoder has been mapped on the FPGA device; it is split into two functional sub- blocks: the first is the de-interleaver,