Error sensitivity data structures and retransmission ... - IEEE Xplore

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003

872

Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging Marco Grangetto, Enrico Magli, Gabriella Olmo Abstract — In this paper we address the problem of JPEG 2000 imaging in a wireless environment. We first define a flexible and efficient data structure for the description of the error sensitivity of different parts of a JPEG 2000 codestream or file format; the data structure is designed in such a way that it can be seamlessly integrated as payload of a JPEG 2000 marker segment or file format box. Moreover, we investigate ARQ policies for robust packet-based JPEG 2000 image transmission over 3G mobile communication systems, and highlight how the proposed data structure can be exploited to improve the end-to-end performance 1. Index Terms — Wireless JPEG 2000, JPWL, ARQ, error

control, robust communications.

wireless

I.

image

transmission,

3G

INTRODUCTION

JPEG 2000 [1] is the most recent lossy and lossless image compression technology developed by the ISO/IEC SC 29 WG 1. With respect to previous standards, and to other compression schemes, JPEG 2000 provides a number of remarkable features, including improved coding efficiency, fine-grain scalability, support for region of interest coding, random codestream access, and error resilience (see [2], [3] and [4] for excellent tutorials). Thanks to its advanced features, JPEG 2000 has plenty of possible applications, ranging from digital cameras to medical imaging, remote sensing, compression and storage of compound documents, as well as compression and delivery of multimedia images and video. Among all possible target applications for JPEG 2000, a key role is played by the wireless imaging applications. As a matter of fact, the wireless technology is gaining increasing importance in the multimedia arena, and wireless imaging is expected to become a breakthrough application for the successful commercial deployment of wireless mobile communication systems, especially in the 3G and beyond scenarios. As a consequence, the wireless applications can represent an opportunity for the widespread adoption of JPEG 2000 technology. However, wireless imaging is known to pose severe technical challenges, due to the fact that the images must be transmitted from the source to the destination through a 1 This work was supported in part by the MIUR (Italian Ministry of Education and Research) grant FIRB “Primo”. The authors are with CERCOM (Center for Multimedia Radio Communications), Dip. di Elettronica, Politecnico di Torino, Corso Duca degli Abruzzi 24 - 10129 Torino - Italy - Ph.: +39-011-5644195 - FAX: +39011-5644099 - E-mail: grangetto(magli,olmo)@polito.it.

Contributed Paper Manuscript received September 24, 2003

wireless transmission medium, which can potentially cause data losses. Hence, wireless image transmission requires a provision for codestream error robustness, in a such a way that unrecovered bit errors or packet losses do not dramatically reduce the quality of the received data. For these reasons, error resilience has played a major role in the development of Part 1, 2, and 3 of the standard, aiming at defining a technology that would include some basic tools to counteract the effect of errors; these tools are briefly reviewed in Sect. II.B. However, although they are quite effective, these tools have not been designed to provide the very high degree of robustness required by some wireless applications. Therefore, for applications in which the transmission channel is very harsh, additional error protection must often be provided, e.g. in terms of additional forward error correction (FEC) coding, automatic repeat request (ARQ), or any other tool that can guarantee a given degree of quality even in presence of errors. The wireless imaging problem is insofar important, that the JPEG 2000 committee has defined a new work item for the standard, namely Part 11 (Wireless applications), also known as JPWL, which has the purpose of addressing the wireless issues relevant to JPEG 2000. In this paper we propose tools for performing robust transmission of JPEG 2000 images over wireless channels and networks by means of ARQ. In particular, this paper contains two main contributions. The first one consists in the definition of a new data structure that allows an encoder to describe the error sensitivity of different parts of the codestream, and to embed this information into the codestream itself. The error sensitivity description can be exploited for several tasks, including: i) optimizing the codestream robustness via unequal error protection, by assigning a higher degree of protection to the most error-sensitive parts; ii) performing intelligent ARQ, by dedicating a higher number of retransmissions, to the most important packets, possibly based on the received error sensitivity information; iii) carrying out optimized video streaming, by prefetching the most important packets into the packet scheduler; iv) optimizing the operation of a rate transceiver, which can look up the error sensitivity information to understand which quality would be delivered at a given rate. This data structure is also extended to the sensitivity description of file formats. The second main contribution of this paper consists in the investigation of the performance of ARQ for JPEG 2000 in the context of 3G communication systems, possibly exploiting the error sensitivity information, in order to optimize the end-to-end quality of the received images. Although ARQ has already been studied for other encoders and video systems, to the authors’ best knowledge an

0098 3063/00 $10.00 © 2003 IEEE

M. Grangetto et al.: Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging

analysis targeted to JPEG 2000 has not been reported yet. It is worth noticing that part of the proposed techniques, and namely the error sensitivity data structure, have been submitted for inclusion in Part 11 of JPEG 2000; however, the reader should take the final standard, and not this paper, as the definitive reference. This paper is organized as follows. In Sect. II we review the state-of-the-art on robust image coding and transmission, and provide a brief overview of the error resilience tools in JPEG 2000. In Sect. III we outline the main requirements for error sensitivity description, whereas in Sect. IV we discuss the information to be included in the data structure, and in Sect. V we define the data structure format. In Sect. VI we investigate the performance of several ARQ strategies for JPEG 2000; finally, in Sect. VII we draw some conclusions and outline possible research developments. II. BACKGROUND In the following we briefly review the state-of-the-art on robust image coding and transmission (Sect. II.A), and provide an overview of error resilience in JPEG 2000 (Sect. II.B). A. Robust image transmission A large body of research work has been recently carried out on efficient protection techniques for reliable image and video delivery. In particular, it is recognized that using a single protection level (i.e. equal error protection) often provides reasonably good quality; however, since the JPEG 2000 codestream is highly scalable, unequal error protection can be carried out to optimize the end-to-end system performance. 1) In-bitstream error resilience Error resilience can be provided by the encoder itself. Practically all multimedia encoders employ in-bitstream markers to allow decoder resynchronization in case of errors. The most recent encoders use reversible variable length codes [5] to perform forward/backward decoding (as in MPEG-4). It has also been shown that other robust entropy coders, e.g. arithmetic codes with a forbidden symbol, can effectively perform error detection and correction [6]. Another popular strategy consists in data partitioning [7], which amounts to place the most error-sensitive parts of the bitstream contiguously, in such a way that they can be more easily protected against errors. 2) Error detection and correction The error detection functionality is crucial in wireless image transmission. Although some encoders do provide some intrinsic error detection capability [6], error detection is often demanded to outer cyclic redundancy codes (CRC) [8]. This requires that the codestream be fragmented into packets, with a proper number of CRC symbols appended to each packet. When the channel or network conditions are very harsh, inbitstream error resilience tools may be unable to provide a sufficiently high degree of protection. In such cases, FEC codes are often employed to provide additional redundancy, to be used for error detection and correction. Reed-Solomon (RS)

873

codes are very popular in this respect, and have found widespread application for erasure channels (see e.g. [9]). Convolutional codes are very useful for bit-error channels [10], and exhibit nice low-complexity features at both the encoder and decoder side. Turbo-codes have been recently gaining increasing popularity due to their excellent error correction performance [11]; in [12] turbo-codes have also been successfully applied to JPEG 2000. It is worth noticing that RS, convolutional codes and turbo-codes can be made rate-adaptive by means of puncturing the output stream [13], thus allowing to carry out unequal error protection by using a single punctured code. In case of unequal error protection, one has to decide how to fragment the codestream, and which codes must be assigned to each fragment. This problem has been addressed by several authors. In [9][14][15] the allocation problem is solved for RS codes to be used on erasure channels; whereas in [16] the allocation is investigated for convolutional codes. The allocation is typically based on the image rate-distortion (R-D) curve. 3) ARQ Another popular error control strategy is represented by ARQ; with respect to encoder-based protection, ARQ provides lower complexity. However, there are cases in which ARQ does not provide benefits, i.e. when the delay requirements are very stringent, or in case of multicasting, in which requests of broadcast retransmission of different parts of the codestream by different users tend to waste the available bandwidth [15]. ARQ and FEC can be used together in hybrid ARQ schemes [15], in which the corrupted bits of a packets can be used jointly with the (possibly corrupted) bits of the retransmitted packet to incrementally correct errors. 4) Other techniques Other robust coding techniques have also been proposed. In layered coding one generates a codestream whose truncation is optimized only at some given points, i.e. in correspondence of the layers. This strategy is also adopted in JPEG 2000, where layers of arbitrary granularity can be created. On the other hand, multiple description coding [18] consists in the generation of two or more correlated descriptions of the same image. In the most popular setting, all descriptions contribute equally to the quality of the received image, and the final quality depends only on the number of received descriptions [9]; unbalanced multiple description schemes have also been recently proposed [19]. A comparison between layered and multiple description coding schemes in several application settings is reported in [20]. 5) Decoder error concealment Although error control is very helpful in counteracting channel errors, unrecovered errors are not rare in the wireless context. Decoder error concealment consists in estimating the missing data from the correctly received ones. It is worth noticing that, while there has been a large body of research on error concealment for DCT-based image and video coders, less work has been done so far for wavelet-based coders. The

874

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003

problem of estimating missing wavelet coefficients has been addressed in [21]; error concealment for JPEG 2000 coded images has been proposed in [22] by estimating lost bits from the statistics of the entropy coder. B. Error resilience in JPEG 2000 The JPEG 2000 standard provides several error resilience tools, which are briefly summarized in the following; the interested reader can refer to [1][3] for a more detailed description. At the entropy coding level, the main error resilience tool is the termination of the arithmetic coder at the end of each coding pass [25]. This allows to detect errors if, after a coding pass, the arithmetic decoder is not in a predictable state. If coupled with context reset after each coding pass, it also allows effective error concealment by skipping the last part of the codestream, which likely contained errors, and continue decoding from the next coding pass; this is much more efficient that stopping decoding upon detection of an error. Moreover, an additional segmentation symbol can be inserted at the end of each coding pass to make error detection more reliable. Another option is to bypass the arithmetic coder; however, while this does provide increased robustness, it significantly decreases coding efficiency. At the codestream level, resynchronization markers can be inserted, which allow the decoder to recognize packets in the codestream even in presence of errors. Moreover, packet headers can be stripped away and grouped in the main header or tile part header to facilitate their protection by means of the techniques described in Sect. II.A. III. ERROR SENSITIVITY REQUIREMENTS In [23] it is shown that RS codes can be effectively employed to protect the most sensitive part of the bitstream, i.e. the headers. However, while protecting the headers ensures that the decoder will not crash (provided that all the errors in the headers have been corrected), it still does not provide any guarantee on the quality of the decoded image. To this purpose, protection of the entropy-coded data must be carried out, so as to make sure that the decoder will be able to decode the most important parts of the codestream and provide a reasonable quality level. In the framework of an image compression standard, it is desirable that the selected protection techniques be compatible with previous parts of the standard to the largest possible extent. As for JPEG 2000, backward compatibility amounts to ensuring that the redundant information, which can be employed either to protect the data at the encoder or to facilitate error recovery at the decoder, be defined in such a way that a decoder compliant with Part 1 (and/or Part 2/3) will not crash upon decoding of the protected codestream. It must be noticed that this requirement does not necessarily imply that decoding must be carried out correctly; rather, it imposes that the “structure” of the protected codestream is backward compatible, by e.g. inserting the additional redundancy information (for example parity symbols of a RS encoder) in a new marker segment that can be skipped by a decoder that

does not exhibit the required advanced decoding features. On a related note, it must be pointed out that a JPEG 2000 Part 1 compliant codestream transmitted over a wireless channel, if affected by errors in the headers, would not guarantee any longer to not crash its decoder; consequently, codestream backward compatibility should be pursued as far as possible, but in the end complete compatibility depends on the correct reception of the codestream headers, which must be guaranteed by suitable error control techniques such as FEC or ARQ. A couple of examples highlighting this issue are provided hereafter. As has been shown in Sect. II.A, unequal error protection can be achieved e.g. by means of selective FEC coding and ARQ. Selective FEC coding, in much the same way as other encoder-based techniques, suffers from the problem that the chosen code type and parameters, or sets thereof, must also be conveyed to the receiver. More importantly, this side information must be assigned the highest priority, since it is required in order to extract the codestream and hence recover the image data. It is not trivial to embed this side information in the codestream itself; this could require the availability of a reliable control channel. To a Part 1/2/3 decoder, the protected codestream would appear as an invalid codestream, potentially leading to a crash. On the contrary, ARQ at the decoder can avoid this problem. The headers can be obtained error-free by e.g. assigning them the maximum number of retransmissions, whereas the remaining retransmission attempts after header reception can be optimally assigned to the other missing packets so as to optimize the end-to-end peak signal-to-noise ratio (PSNR). If the headers are not correctly received after all available retransmission attempts, the image is considered lost; this avoid decoding a codestream whose headers are potentially corrupted. As has been seen, error protection involves several different kinds of information that have to be conveyed to the receiver. These kinds of information can be classified as follows. − − −

The protected representation of the data (e.g. data plus parity symbols, convolutionally encoded data, interleaved data, and so on). Side information related to the type of protection employed and the related parameters. The codestream error sensitivity information.

In this paper we propose a data structure suitable for describing information of the last type, i.e. error sensitivity information of JPEG 2000 codestreams, which can be embedded in the codestream itself in a backward-compatible way. This can be obtained by defining a marker and a related marker segment that contain the proposed data structure. A Part 1/2/3 compliant decoder, not knowing the meaning of the new marker, will simply skip it and continue decoding the codestream. Conversely, an advanced decoder can take advantage of the knowledge of the error sensitivity information in order to optimize its ARQ policy.

M. Grangetto et al.: Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging

IV. ERROR SENSITIVITY DESCRIPTOR Error sensitivity information must provide a measure of how much different parts of the codestream are sensitive to errors, or equivalently the effect of losing each part on the quality of the decoded image. In the following the data structure containing such information will be referred to as Error Sensitivity Descriptor (ESD). The basic building block of the JPEG 2000 codestream is a packet. In some cases (see e.g. [24]) it can be useful to provide a sensitivity measure for every packet in the codestream; this provides indeed a very fine sensitivity description. This case will be referred to in the following as packet mode. However, there are reasons why other indexing schemes would also be useful. For example, since compressed images can consist of a very high number of packets, providing sensitivity information for each packet may lead to a large overhead. Moreover, the complexity of the algorithm that unequally allocates channel codes to different portions of the codestream will increase with the number of sensitivity levels; for this reason, and given the limited granularity of available channel codes, it is likely that in a practical setting sensitivity would be provided at a reasonably coarse scale, thus partitioning the codestream in a low number of portions. As a consequence, it is important to foresee a second indexing method, which is not tailored to the JPEG 2000 codestream structure, but rather based on byte ranges. This case will be referred to in the following as byte-range mode. A. Sensitivity information 1) Meaning of sensitivity values Sensitivity information shall be expressed as a two byte unsigned integer number ranging from 0 to 216-1. Although it is felt that one single byte (i.e. 256 different sensitivity values) is sufficient for most applications, it is certainly convenient that each sensitivity description record be aligned on a 4 byte boundary. As will be seen shortly, in byte-range mode 3+3 bytes are required to specify the start and the end byte of a segment; consequently, it is convenient to use two bytes for the sensitivity description so as to have alignment on 4-byte boundaries. In packet mode indexing is not necessary, so that only the sensitivity values are stored; two bytes can be used to provide the alignment for each couple of values. It is intended that the lowest values of sensitivity shall be assigned to the “most important” parts of the codestream. The value 0 shall be exclusively reserved for headers, and the value 216-1 for dummy sensitivity records (see Sect. (V.C.5)). All the other values shall represent the relative importance of the considered portion of the codestream, in the [1,216-2] range, with 1 denoting the highest level of importance. 2) Default mode As for the computation of the sensitivity information, two cases can be envisaged. The first case occurs when the encoder generates this information at the very time of image coding; in this case the rate allocator can provide the sensitivity information in terms of mean-squared error (MSE), PSNR or other metrics, which can be used to fill the error sensitivity

875

data structure. The second case occurs when the following two conditions are met: 1) one wants to transmit an already encoded image for which the sensitivity information was not generated at the time of encoding, and 2) it is not possible to re-generate this information, e.g. by decoding at several different rates and computing the error metrics. In this case, since the rate allocation is made in such a way that the codestream quality is progressive, a default sensitivity mode will be signaled by switching the relevant flag in the ESD data structure (see the PESD parameter described in Sect. V). In the default mode, no explicit sensitivity information is signaled, but it is suggested that the decoder looks up the main header to determine the codestream progression type, and uses that information as best as it can. For example, if the default mode is signaled and the codestream has been formed in quality progressive mode, the decoder can assume that packets will have decreasing importance. This would be equivalent to saying that the first packet has sensitivity one, the second one has sensitivity two, and so on. On the other hand, if the codestream is in component-progressive mode, the decoder may assume that the first packet of each component has sensitivity one, the second one has sensitivity two, and so on. Notice that these conditions are signaled implicitly by switching on the default mode; however, it must be pointed out that it is optional for the decoder to exploit this information. 3) Relative and absolute sensitivity In general, the ESD data structure shall contain relative sensitivity information for each part of the codestream. With the term “relative” it is meant that, in principle, this information will not be tied to any specific error metric such as MSE, PSNR, or absolute peak error. However, since absolute metrics may be useful in some applications, provisions are made in the syntax to relate the sensitivity information to an error metric. In particular, the MSE, PSNR and absolute peak error between decoded and original image are explicitly considered, whereas the syntax can be extended to signal other (possibly custom) error metrics. In particular, an “absolute sensitivity” mode is foreseen, in which a mapping from the relative values provided in the data records of the ESD data structure to metric values is provided by means of the OG parameter. This parameter consists of an (offset,gain) couple so that, for each relative sensitivity value S, the metric specified in PESD is computed as M = S ⋅ gain + offset . The OG parameter consists of four bytes; the first couple of bytes represent the offset, whereas the second one the gain. Either number is represented as a real number in pseudo floatingpoint notation. The pseudo floating-point format is similar to that used for signalling quantization step sizes in Part 1 [1], and is defined as follows. Each 16 bit number contains the sign (1 bit), exponent (5 bits) and mantissa (10 bits) of the offset (resp. gain) parameter. In particular, the true value V of the offset (respectively gain) parameter is given by the following formula:

876

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003

µ   V = ±2 ε 1 + 10   2  where the sign is positive if the first bit of the 16-bit number is 1 and negative otherwise; ε is the decimal number obtained from the next five bits of the parameter, and µ the decimal number obtained from the remaining 10 bits. 4) Provision for rate-distortion models Finally, it is worth noticing that, in order to decrease the complexity, models of the rate-distortion (R-D) curve can be employed instead of computing the true R-D curve (see e.g. [17]). R-D models are described by a certain number of parameters, and few points of the true R-D curve are necessary to match the parameters to the R-D curve of the current image. The ESD data structure may provide additional syntax to embed these parameters in the codestream, so that the decoder can build the R-D model and compute suitable error sensitivity metrics based on it. In this paper we do not explicitly address the use of R-D models within the ESD marker segment, though it is easy to generalize its structure in order to accommodate this case. B. Packet mode In the packet mode, the sensitivity description is based on packets. The description corresponds to one data record for each packet in the codestream; each record consists of two bytes, which represent the sensitivity value for each packet. Notice that, since the number of packets is known a priori, explicit packet numbering is not necessary; instead, it is assumed that the records appear in the order specified by the packet numbering of the SOP marker (see [1], Annex A.8.1). It is worth noticing that, since the main and tile-part headers are not contained in any JPEG 2000 packet, in packet mode the sensitivity value 0 should not be used. C. Byte-range mode In byte-range mode, the sensitivity information is not any more constrained by the packet structure of the JPEG 2000 codestream; instead, sensitivity can be arbitrarily specified for any part of the codestream. The description corresponds to one data record for each unit in the codestream. Each data record consists of 8 bytes. The first three bytes contain the start byte of the data unit in the codestream; the next three bytes contain the end byte of the data unit (byte numbering starts from one, i.e. the SOC marker has range from 1 to 2). The remaining two bytes contain the sensitivity value for the data unit. The number of data units can be derived from the length parameter after subtracting the length of the PESD parameter, and of the OG parameter if present. V. DATA STRUCTURE DEFINITION In this section we define an ESD data structure which contains the sensitivity information for a given codestream or tile. When using integer numbers, we always assume that these numbers are stored in the data structure in big endian order, unless otherwise explicitly stated.

A. Format and possible use within a marker segment In order to ensure backward compatibility, the ESD data structure shall be included in the standard JPEG 2000 codestream headers, i.e. in the main header and/or in the tile part headers. In the following we define an ESD marker segment following the rules specified in [1]. In particular, the numbering and format of this marker shall adhere to the conventions in [1], i.e. the marker is two bytes long, and its first byte value is 0xFF; the second byte specifies the marker use and can take any value in the range 0x01 to 0xFE, apart from those already used by the ITU-T Rec. T.81 | ISO/IEC 10918-1 and ITU-T Rec. T.84 | ISO/IEC 10918-3 (see [1]). Subsequently, the next two bytes must correspond to an unsigned big endian integer value that denotes the length in bytes of the marker parameters (including two bytes of this length parameter but not including the two bytes of the marker itself). If the ESD data structure is in the main header, it shall be placed after the SOC and SIZ marker segments, as dictated by [1], and after the EPB marker segment2, as dictated by [23]. If it is in a tile-part header, it shall be placed after the SOT and EPB markers and before the SOD marker. If an ESD data structure is included both in the main header and in the tilepart headers, and they both describe sensitivity of the same or overlapping portions of the codestream (e.g., in the byte range mode), it is intended that the information in the tile-part header overrides that in the main header as to the error sensitivity of the overlapping portions of the description. In the byte-range mode, it is allowed for an ESD data structure to describe the sensitivity of codestream portions further away than the EOC marker. Similarly, in packet mode it is allowed to anticipate sensitivity values for packets in the next codestream by addressing packets beyond the EOC marker (e.g., in the next codestream); notice that, since the headers are not contained in any packet, in packet mode it is only possible to describe the sensitivity of the entropy-coded data in the codestream. This “sensitivity anticipation” can be useful for video transmission, in which, for each frame, the ESD information can be used to infer where the headers of the codestream for the next frame are placed in the data stream. Since it is possible that a decoder has two different sensitivity values for a given codestream part (e.g., one from the previous codestream and one from the current one), the sensitivity information contained in the current codestream will override the information in other codestreams referring to portions of the current codestream. B. ESD data structure description Figure 1 describes the syntax proposed for the ESD data structure. It consists of the following fields: − ESD: code identifying the data structure. − LESD: length of the data structure in bytes. 2 The EPB marker segment is a new marker which is being defined within Part 11, with the aim of containing parity symbols for protection of the codestream headers.

M. Grangetto et al.: Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging

ESD LESD PESD

−

PESD: field describing the usage of the data structure. OG: (offset,gain) parameter for conversion to absolute sensitivity values. ESD data: Record of parameters related to error sensitivity.

OG

− −

ESD data

Figure 1 - Error Sensitivity Description (ESD) syntax.

C. Parameter description 1) ESD parameter This parameter is an unsigned integer of length 2 bytes, which contains a unique hexadecimal code identifying the beginning of the data structure. For example, it can be a valid JPEG 2000 marker dedicated to the error sensitivity description. It is actually foreseen that Part 11 of JPEG 2000 will define an ESD marker with code 0xFF98. 2) LESD parameter This parameter is an unsigned integer of length 2 bytes, which can take values in the [7, (216-1)] range, and specifies the length in byte of the ESD data structure excluding the ESD parameter and including the LESD parameter itself. 3) PESD parameter This parameter is a 4-byte long binary number (in the 0 — 232-1 range) specifying the usage of the data structure. The meaning of the bits in the PESD field is specified below. Bit 1: the first bit is always set to one. Bit 2: set to 0 or 1 if the packet mode or the byte-range mode are used respectively. Bit 3: it is set to 1 if the default mode is used; otherwise it is set to zero. If the default mode is used, the OG paramer and the ESD data field are not present. Bit 4-6: provide information on the type of error sensitivity description. 000 means relative error sensitivity, in which case the OG parameter is not present. Other values indicate “absolute sensitivity” modes, for which the OG parameter can be used to translate the relative sensitivity values into absolute ones. In particular, 001 means that the absolute metric is MSE, 010 means PSNR, 011 to absolute peak error, 100 means some custom metric which is agreed upon between encoder and decoder, but is not signaled in this marker segment. The other values are reserved. Bits 7-32: reserved for future use. 4) OG parameter This parameter is present only if the default mode is not used, and if one of the “absolute sensitivity” modes is enabled in PESD. OG is a couple of 16 bit pseudo floating-point values. The first value is the “offset” coefficient, whereas the second one is the “gain” coefficient. The pseudo floating-point representation, as well as the translation of relative sensitivity values into absolute ones, are as described in Sect. (IV.A.3).

877

TABLE I AVERAGE PSNR FOR MOTHER & DAUGHTER (QCIF FORMAT, 10 FPS) PLR

M

0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 0.15 0.15 0.15 0.15 0.20 0.20 0.20 0.20

0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞

64 kbit/s 26.74 29.03 29.10 29.10 25.60 28.40 28.78 28.82 24.76 27.58 28.42 28.56 24.11 26.90 27.94 28.29

PSNR (dB) 128 kbit/s 384 kbit/s 27.37 27.64 32.07 38.26 32.30 40.40 32.30 40.56 25.81 25.85 30.68 33.42 31.87 38.68 32.01 39.97 24.84 24.86 29.13 30.36 31.26 36.33 31.72 39.36 24.15 24.15 27.93 28.50 30.36 33.56 31.42 38.76

512 kbit/s 27.64 39.71 43.57 43.93 25.85 33.72 40.68 43.18 24.86 30.45 37.42 42.49 24.15 28.53 34.02 41.68

TABLE II AVERAGE PSNR FOR MOTHER & DAUGHTER (QCIF FORMAT, 15 FPS) PLR

M

0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 0.15 0.15 0.15 0.15 0.20 0.20 0.20 0.20

0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞

64 kbit/s 25.88 27.07 27.10 27.10 25.15 26.76 26.91 26.93 24.51 26.30 26.71 26.78 23.97 25.91 26.44 26.60

PSNR (dB) 128 kbit/s 384 kbit/s 27.10 27.61 30.29 35.94 30.40 36.88 30.40 36.93 25.74 25.85 29.38 32.66 30.00 35.84 30.07 36.37 24.82 24.86 28.27 30.08 29.53 34.46 29.76 35.92 24.13 24.15 27.38 28.40 28.90 32.58 29.44 35.51

512 kbit/s 27.64 37.52 39.15 39.26 25.85 32.22 37.76 38.72 24.86 30.29 35.79 38.22 24.15 28.48 33.32 37.76

5) ESD data field This data field has variable length. It is present only if the default mode is disabled in PESD. The length of the ESD field can be derived by LESD, subtracting the bytes used by PESD and by the OG parameter if present. It is suggested to not derive its length from the number of packets in the codestream or from other prior knowledge on the codestream size, since it is possible that the ESD segment addresses more data than are contained in the codestream, e.g., in order to anticipate information in the next codestream. The ESD field can have two different syntaxes according to whether the packet mode or the byte-range mode is selected. Packet mode: in this case each record is two bytes long, and consecutively contains the sensitivity information (unsigned integer) for all packets in the codestream or tile. In order to provide alignment on 4-byte boundaries, if an odd number of packets is present, a dummy sensitivity value will be inserted as last record, with value 216-1. Byte-range mode: in this mode, the length of each record is 8 bytes. The first three bytes contain the start byte of a data unit in the codestream; the next three bytes contain the end

878

byte of the data unit (byte numbering starts from one). The remaining two bytes contain the sensitivity value for the data unit (unsigned integer). It is assumed that the first byte of the codestream shall be byte one. D. Error sensitivity description for file formats In some applications it is more useful to transmit a file format instead of a raw codestream, since a file format provides the syntax for embedding ancillary information about the media. The JPEG 2000 family of standards provides a number of file formats, such as the JP2, JPX, MJ2 and JPM, which are designed to match plenty of applications ranging from image and video transmission to remote sensing. Of particular interest for the wireless applications is the MJ2 file format for compressed video [27], which can contain a sequence of codestreams interleaved with other video-related information. It is clear that, in the wireless applications, all or part of this ancillary information may be necessary in order to be able to decode the video sequence. As a consequence, it is very important to devise a way to describe error sensitivity for a file format. The ESD marker segment previously described can serve this purpose with minor modifications. In particular, we define a new box, which can be used in any JPEG 2000 file format, and can contain sensitivity information for the whole file format. The box, named Error Sensitivity Descriptor box (ASCII code ‘esdb’), follows the positioning and syntax conventions defined by the relevant file format in which it is inserted; as an example, for the JP2 format this means that the ‘esdb’ box shall follow the JPEG 2000 signature box and the File Type box. It shall consist of a 4-byte LBox parameter specifying the box length, followed by a 4-byte TBox parameter containing the ‘esdb’ code, followed by an optional XLBox extended box length parameter, followed by the DBox box content field. The DBox field shall contain a data structure which is defined exactly in the same way as the ESD marker segment, except for the fact that the ESD and LESD fields shall not be present. The information contained in a box shall be defined and employed in the same way as the ESD marker segment, with the following modifications. − The parameter to be used for computation of the length of the ESD data field shall be the LBox and not LESD. − In the definition of sensitivity values, those parts of the file which do not contain codestreams shall be regarded as headers (i.e. they shall be given sensitivity equal to 0). − It is possible that an ‘esdb’ box outside the codestream and an ESD marker segment inside the codestream provide sensitivity values related to the same codestream portion. If this happens, the information in the ESD marker segment shall override that in the ‘esdb’ box. − Since some files may be very large, especially in case of video sequences, there is a problem of addressing such files in byte-range mode and packet mode. It is up

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003 TABLE III AVERAGE PSNR FOR MISS AMERICA (CIF FORMAT, 5 FPS) PLR

M

0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 0.15 0.15 0.15 0.15 0.20 0.20 0.20 0.20

0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞

64 kbit/s 21.46 33.00 35.23 35.23 19.58 28.66 33.75 34.97 17.94 24.68 31.26 34.73 16.85 22.78 28.04 34.47

PSNR (dB) 128 kbit/s 384 kbit/s 21.47 21.48 34.40 35.56 38.08 42.16 38.10 42.29 19.58 19.58 29.07 29.28 35.67 37.60 37.86 42.11 17.94 17.94 24.80 24.85 32.24 32.99 37.60 41.88 16.85 16.85 22.83 22.84 28.45 28.70 37.36 41.59

512 kbit/s 21.48 35.75 43.36 43.55 19.58 29.30 37.95 43.27 17.94 24.85 33.09 43.04 16.85 22.85 28.73 42.75

TABLE IV AVERAGE PSNR FOR MISS AMERICA (CIF FORMAT, 10 FPS) PLR

M

0.05 0.05 0.05 0.05 0.10 0.10 0.10 0.10 0.15 0.15 0.15 0.15 0.20 0.20 0.20 0.20

0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞ 0 1 2 ∞

−

64 kbit/s 21.38 31.06 32.30 32.30 19.55 27.89 31.39 32.03 17.93 24.39 29.75 31.81 16.85 22.62 27.26 31.57

PSNR (dB) 128 kbit/s 384 kbit/s 21.46 21.48 33.00 34.94 35.23 39.63 35.23 39.69 19.58 19.58 28.66 29.19 33.75 36.56 34.97 39.53 17.94 17.94 24.68 24.83 31.26 32.64 34.73 39.35 16.85 16.85 22.78 22.84 28.04 28.59 34.47 39.11

512 kbit/s 21.48 35.21 40.58 40.66 19.58 29.23 36.98 40.46 17.94 24.84 32.79 40.24 16.85 22.84 28.64 40.02

to the encoder to ensure that the addressed byte ranges in a single ‘esdb’ box do not overflow when trying to address data further away off the codestream next to the ‘esdb’ box. Unlike the raw codestream, the ‘esdb’ box may be located outside a codestream, and there can be several ones in the same file format. For each ‘esdb’ box, in packet mode the packet numbering shall refer to the JPEG 2000 packets of the codestreams following the ‘esdb’ box; in byte-range mode the byte ranges shall be referred to the first byte of the ‘esdb’ box itself (byte one).

VI. ARQ STRATEGIES FOR ROBUST JPEG 2000 TRANSMISSION OVER MOBILE COMMUNICATION SYSTEMS In the following we investigate the performance of ARQbased schemes for robust JPEG 2000 wireless image transmission, exploiting the error sensitivity information previously described.

M. Grangetto et al.: Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging

A. Network model and video application As target application, we consider image and video communications in 3G mobile cellular systems using JPEG 2000 or Motion JPEG 2000 [27]. In particular, the network model is depicted in Fig. 2. The video data are conveyed from a wired network to a base station, which forwards them to the end-user via a wireless link. The base station may contain a proxy server as in [15], which parses the incoming packets in a smart way, so as to optimally handle retransmissions to the end-user. The video data can be generated e.g. by a streaming video server, or by another mobile terminal in case of conversational application. In this paper we consider the downlink between base station and terminal. This corresponds to the whole wireless link in the streaming video case, whereas it only represents the base station to terminal air interface in case of conversational applications; in this latter case, the proxy server should also ask for retransmissions from the video source terminal.

879

Figure 3 – MOTHER & DAUGHTER QCIF picture.

Wired network

Terminal

JPEG 2000 decoder

Base station

Figure 2 - Wireless network model.

We consider video transmission on a frame-by-frame basis. Grayscale images in CIF (Common Intermediate File Format) and QCIF (Quarter CIF) format have been considered (i.e., 352x288 and 176x144 pixels respectively). Channel data rates of 64, 128, 384 and 512 kbit/s have been taken into account; these rates can be easily accommodated by 2.5G/3G mobile communication system, and can provide reasonably good video quality. Packets of 80 bytes each have been considered at the RLC (radio link control) layer, and the RLC acknowledged mode has been used to carry out ARQ at the link layer. This setting is typical of image and video transmission over a UMTS radio link for conversational applications [26]. The video frame-rate has been selected so as to provide a reasonably good quality to the end user (i.e. temporally smooth playback with sufficiently high PSNR) with respect to the available bandwidth and the JPEG 2000 compression capability. This has led to the selection of 10 and 15 frames/s for QCIF pictures, and 5 and 10 frames/s for higher resolution CIF pictures; these corresponds to source bitrates from 0.2 to 1.5 bit-per-pixel (bpp). The data rate, packet size and frame rate constrain the maximum number N of packet transmission attempts available for each picture (including possible retransmissions). B. ARQ strategies We have considered the following ARQ policy that exploits

Figure 4 – MISS AMERICA CIF picture.

the error sensitivity information contained in the ESD data structure. We assume that a proxy server at the base station intelligently manages retransmissions towards the end-user terminal. The intelligence lies in the capability of the proxy server to parse a JPEG 2000 or Motion JPEG 2000 codestream and i) identify the picture number and codestream start and end bytes, ii) identify the main and tile-part headers, which are vital for correct decoding, iii) extract the ESD data structure. The ESD can be employed to adapt the retransmission strategy according to the error sensitivity of the different packets, and/or to carry out rate adaptation with exact knowledge of the delivered quality for each possible rate by means e.g. of the absolute sensitivity information. Notice that, if a codestream employs the layer-based progression order to optimize the expected PSNR at the receiver, a good packet retransmission order is one that follows increasing packet numbering. Although this approach is not guaranteed to be optimal in terms of quality, because it does not take into account decoder error concealment by means of the JPEG 2000 resilience tools, in our experiments it has turned out to provide excellent results. As a consequence, if the proxy server

880

must only carry out retransmissions, the default mode in the ESD structure usually provides enough information. Nevertheless, it must be noticed that absolute sensitivity information is crucial to enable accurate quality control for rate adaptation, or hybrid ARQ and FEC [9][15]. Based on these remarks, we have considered an ARQ system which processes packets sequentially; in particular, each unacknowledged packet is assigned a maximum number M of retransmissions before the following packet is processed. Transmission is stopped upon reaching N transmission attempts, corresponding to the available bandwidth for the current picture. This is a customary approach when the proxy or base station is not aware of the content of each packet, and attempts to provide the same delay for each packet. As a consequence, some packets may be dropped, and the JPEG 2000 decoder must employ its error resilience capability to conceal the effect of losses. The case for M=0 corresponds to no error protection via ARQ at all; consequently, all the available bandwidth is used for coded source data. Moreover, we have investigated the behavior of an ARQ system which only enforces a maximum number of retransmissions per picture N, but does not constrain the number of retransmissions for each packet; this is equivalent to the previous ARQ strategy in the limit of M=∞, constrained to the maximum number N. Such system is expected to provide improved performance, since it recognizes that the delay constraint can be enforced on a per-picture basis, and hence employs the maximum number of leftover retransmission for each packet, thus guaranteeing that the first and most important packets in the codestream are assigned a higher number of attempts. C. Experimental results In the following we report experimental results on the transmission of JPEG 2000 coded pictures according to the scenario described above. The MOTHER & DAUGHTER (QCIF) and MISS AMERICA (CIF) pictures have been used to evaluate the performance of the considered ARQ schemes; these pictures are shown in Fig. 3 and 4 respectively. The JPEG 2000 encoder provides error resilience by using the SOP and EPH markers [1], resetting the contexts and terminating the arithmetic coder at each coding pass, and inserting the segmentation symbol. The layer-progressive order is employed using 50 layers. It is assumed that the headers are received error-free by the terminal. This can be obtained by either using FEC codes to protect the headers as in [23], or by always devoting M≤ N retransmissions to the packets that contain the main and tile-part headers. The former strategy has been used in our experiments. Expected PSNR (computed by linearly averaging the MSE) has been used as quality metric, averaging the results of a Monte Carlo simulation over 1000 picture transmissions for each test condition. Random packet loss rates (PLR) ranging from 5% to 20% have been selected as representative of several wireless scenarios including mobile communications.

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003

In Tab. I the results for the MOTHER & DAUGHTER picture at 10 frames/s are reported for several PLRs and data rates. As can be seen, if no retransmissions at all are used (M=0), a poor average quality is obtained also at PLRs as low as 5%, even though the headers are received error-free. This points out that some form of error control is required in order to provide the user with an acceptable quality level. In fact, with M=0 the received quality is dominated by the position of the first lost packet, and this justifies the fact that PSNR is little dependent on the data rate. Employing one or two retransmissions yields a large PSNR gain, showing the suitability of the ARQ approach for this application. It is worth noticing that the M=∞ approach always provides the best results, highlighting the validity of the proposed approach. However, at PLRs as low as 5%, two retransmissions yield quality very close or equal to the maximum values. On the other hand, for very harsh conditions, the gain of the M=∞ approach becomes significant, pointing out that limiting the number of retransmissions per packet is a suboptimal strategy. This is due to the fact that the probability of losing one of the first and most critical packets is high, thus vanishing the quality improvements conveyed by the next received packets. The performance improvement of the M=∞ approach with respect to M=2 increases with higher channel data rates, since this allows the proxy server to better manage the retransmission of the most critical packets. Tab. II reports the results of the same experiment, encoding the picture at 15 frames/s. Apart from the PSNR decrease due to the increased frame rate (and hence smoother playback), similar comments can be made as for the quality achieved by the considered ARQ strategies at various PLRs and data rates. Tab. III and IV show the results of the same experiment for a CIF image encoded at 5 and 10 frames/s respectively. It can be observed that the same considerations as before still hold for this higher resolution image format. As a consequence, it can be seen that the proposed M=∞ strategy consistently provides the best results for this kind of application. VII. DISCUSSION AND CONCLUSIONS As has been shown in Sect. VI, error sensitivity information can be exploited by a proxy server at the base station to optimize the retransmission management. In particular, it has been found that the M=∞ strategy, which does not constrain the number of retransmission attempts for each packet, but rather for each frame, provides improved performance. This is mainly due to the fact that this strategy privileges the retransmission of the first packets of the codestream, achieving a good trade-off between channel rate and delivered quality. If the codestream has been encoded in layer-progressive order, the default mode for error sensitivity description turns out to be adequate. On the other hand, it is worth noticing that, if other progression orders are employed, requesting packets sequentially may not be the best approach. As an example,

M. Grangetto et al.: Error Sensitivity Data Structures and Retransmission Strategies for Robust JPEG 2000 Wireless Imaging 32 30

[11]

28

[12]

PSNR (dB)

26 24

[13]

22

[14]

20 18

[15]

16 14 12

0

0.25

0.5

0.75 Rate (bpp)

1

1.25

1.5

Figure 5 – Typical R-D curve of an RGB picture in componentprogressive order (the plot refers to the BOATS image).

consider the transmission of a color RGB picture in componentwise progression order, i.e. the codestream contains first all data related to the R, then to the G and B components. A typical R-D curve of such a codestream is shown in Fig. 5 for the BOATS image (PSNR is referred to each component separately). It can be clearly seen that, in this case, the codestream portions that mostly contribute to the received quality are not placed at the beginning of the codestream, but at the beginning of each component. As a consequence, the M=∞ approach could be used by retransmitting packets not in sequential order, but according to the information in the ESD data structure. REFERENCES [1]

ISO/IEC JTC 1 SC 29 WG 1 15444-1 “JPEG 2000 Part 1 – Core coding system”. [2] D.S. Taubman, M.W. Marcellin, “JPEG 2000: standard for interactive imaging,” Proceedings of the IEEE, vol. 90, n. 8, pp.1336-1357, Aug. 2002. [3] A. Skodras, C. Christopoulos, T. Ebrahimi, “The JPEG 2000 still image compression standard,” IEEE Signal Processing Magazine, pp.36-58, Sept. 2001. [4] C. Christopoulos, A. Skodras, T. Ebrahimi, “The JPEG2000 still image coding system: an overview,” IEEE Transactions on Consumer Electronics, vol. 46, n. 4, pp. 1103-1127, Nov. 2000. [5] I. Moccagatta, S. Soudagar, J. Liang, H. Chen, “Error-resilient coding in JPEG-2000 and MPEG-4,” IEEE Journal on Selected Areas in Communications, vol. 18, n. 6, pp. 899-914, Jun. 2000. [6] M. Grangetto, E. Magli, G. Olmo, “Robust video transmission over error-prone channels via error correcting arithmetic codes,” IEEE Communications Letters, to appear 2003. [7] J. Wen, J. Villasenor, H.D. Shin, “Proposal for error resilience in H.26L,” ITU-T documentation Q15-C-36, Eibsee, Germany, 1997. [8] P.G. Sherwood, K. Zeger, “Progressive image coding for noisy channels,” IEEE Signal Processing Letters, vol. 4, n. 7, pp. 189-191, Jul. 1997. [9] A.E. Mohr, R.E. Ladner, E.A. Riskin, “Unequal loss protection: graceful degradation of image quality over packet erasure channels through forward error correction,” IEEE Journal on Selected Areas in Communications, vol. 18, n. 6, pp. 819-828, Jun. 2000. [10] J. Hagenauer, N. Seshadri, C.-E.W. Sundberg, “The performance of rate-compatible punctured convolutional codes for digital mobile radio,”

[16] [17]

[18]

[19] [20] [21] [22] [23] [24] [25] [26] [27]

881

IEEE Transactions on Communications, vol. 38, n. 7, pp. 966-980, Jul. 1990. C. Berrou, A. Glavieux, “Near-optimum error correcting coding and decoding: turbo-codes,” IEEE Transactions on Communications, vol. 44, n. 10, pp. 1261-1271, Oct. 1996. B.A. Banister, B. Belzer, T.R. Fischer, “Robust image transmission using JPEG2000 and turbo-codes,” IEEE Signal Processing Letters, vol. 9, n. 4, pp. 117-119, Apr. 2002. J. Hagenauer, “Rate-compatible punctured convolutional codes (RCPC codes) and their applications,” IEEE Transactions on Communications, vol. 36, n. 4, pp. 389-400, Apr. 1988. R. Puri, K.W. Lee, K. Ramchandran, B. Bharghavan, “An integrated source transcoding and congestion control paradigm for video streaming in the Internet,” IEEE Transactions on Multimedia, vol. 3, n. 1, pp. 1832, Mar. 2001. A. Majumdar, D. Grobe Sachs, I.V. Kosintsev, K. Ramchandran, M.M. Yeung, “Multicast and unicast real-time video streaming over wireless LANs,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 12, n. 6, pp. 524-534, Jun. 2002. V. Chande, N. Farvardin, “Progressive transmission of images over memoryless noisy channels,” IEEE Journal on Selected Areas in Communications, vol. 18, n. 6, pp. 850-860, Jun. 2000. Z. He, Y.K. Kim, S. Mitra, “Low-delay rate control for DCT video coding via ρ-domain source modeling,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 11, n. 8, pp. 928-940, Aug. 2001. A.R. Reibman, H. Jafarkhani, Y. Wang, M.T. Orchard, R. Puri, “Multiple-description video coding using motion-compensated temporal prediction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 12, n. 3, pp. 193-204, Mar. 2002. J.G. Apostolopoulos, S.J. Wee, “Unbalanced multiple description video communication using path diversity,” Proc. of ICIP 2001 (IEEE International Conference on Image Processing). Y.C. Lee, J. Kim, Y. Altunbasak, R.M. Mersereau, “Layered coded vs. multiple description coded video over error-prone networks,” Signal Processing: Image Communication, vol. 18, pp. 337-356, 2003. S.S. Hemami, R.M. Gray, “Subband-coded image reconstruction for lossy packet networks,” IEEE Transactions on Image Processing, vol. 6, n. 4, pp. 523-539, Apr. 1997. K. Djafarian, S. Berrada, M. Shima, K. Oehler, “IMAP JPEG2000 decoder error concealment module enables streamed Motion JPEG2000 in 3G wireless networks,” submitted to IEEE ICASSP 2003. D. Nicholson, C. Lamy, C. Poulliat, X. Naturel, “Backward Compatible Header Error Protection in a JPEG 2000 codestream”, also WG1N2851. E. Edwards, S. Futemma, E. Itakura, N. Tomita, A. Leung, T. Fukuhara, “RTP payload format for JPEG 2000 video streams”, also WG1N2768. D. Taubman, “High performance scalable image compression with EBCOT,” IEEE Transactions on Image Processing, vol. 9, n. 7, pp. 1158-1170, Jul. 2000. ITU-T VCEG-N37, “Common Test Conditions for RTP/IP over 3GPP/3GPP2 – Software and Amendments”, VCEG (SG16/Q6), XIV Meeting, Santa Barbara, Sept. 2001. ISO/IEC JTC 1 SC 29 WG 1 15444-3 “Part 3 – Motion JPEG 2000”.

Marco Grangetto received the “summa cum laude” degree and the Ph.D. degree in Electrical Engineering at Politecnico di Torino in 1999 and 2003 respectively. His research interests are in the field of digital signal processing and multimedia communications. In particular, he is working at the development of robust and efficient lossy and lossless image and video encoders based on wavelet transforms. Moreover, he is addressing the application of joint source and channel coding to the design of reliable multimedia delivery systems for wireless lossy packet networking. He was awarded the premio optime by “Unione industriale” di Torino in Sept. 2000, and a Fulbright grant in 2001 for a research period at the Center for Wireless Communications (CWC) at UCSD.

882 Enrico Magli received the degree in Electronics Engineering in 1997, and the Ph.D. degree in Electrical Engineering in 2001, from Politecnico di Torino, Turin, Italy. He is currently an Assistant Professor at the same university, and is affiliated with the CERCOM (Center for Multimedia Radio Communications) research center. His research interests are in the field of robust image and video coding for wireless applications, compression of remote sensing images, and digital watermarking. From Mar. to Aug. 2000 he was a visiting researcher at the Signal Processing Laboratory of the Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland. He has coauthored more than 70 scientific papers in international journals and conferences. He is currently a member of the Data Archiving and Distribution (DAD) Technical Committee of the IEEE Geoscience and Remote Sensing Society, and a contributor to the ISO activities on JPEG 2000 (Part 11, wireless applications). He has been member of the Technical Program Committee and session chair for several international conferences, including IEEE International Conference on Multimedia and Expo (ICME) 2002 and 2003, IEEE International Conference on Image Processing (ICIP) 2003.

IEEE Transactions on Consumer Electronics, Vol. 49, No. 4, NOVEMBER 2003 Gabriella Olmo received the Laurea Degree (cum laude) and the PhD in Electronic Engineering at Politecnico di Torino in 1986 and 1992 respectively. From 1986 to 1988 she was researcher with CSELT (Centro Studi e Laboratori in Telecomunicazioni), Turin, working on network management, non hierarchical models and dynamic routing. From 1991, she has been Assistant Professor at Politecnico di Torino, where she is member of the Telecommunications group and head of the Image Processing Lab; from 2002 she is Associate professor at the same University. Her main recent interests are in the field of wavelets, remote sensing, image and video coding, resilient multimedia transmission, joint source-channel coding, stratospheric platforms. She has joined several national and international research programs under contracts by Inmarsat, ESA (European Space Agency), ASI (Italian Space Agency), European Community. She has coauthored more than 100 papers in international scientific journals and conference proceedings.