While network data transmission was traditionally accomplished via routing, network coding (NC) broke this rule by allowing network nodes to perform linear combinations of the upcoming data packets. Network operations are performed in a specific Galois field of fixed size q. Decoding only involves a Gaussian elimination with the received network-coded packets. However, in practical wireless environments, NC might be susceptible to transmission errors caused by noise, fading, or interference. This drawback is quite problematic for real-time applications, such as multimediacontent delivery, where timing constraints may lead to the reception of an insufficient number of packets and consequently to difficulties in decoding the transmitted sources. At best, some packets can be recovered, while in the worst case, the receiver is unable to recover any of the transmitted packets.In this thesis, we propose joint source-network coding and decoding schemes in the purpose of providing an approximate reconstruction of the source in situations where perfect decoding is not possible. The main motivation comes from the fact that source redundancy can be exploited at the decoder in order to estimate the transmitted packets, even when some of them are missing. The redundancy can be either natural, i.e, already existing, or artificial, i.e, externally introduced.Regarding artificial redundancy, we choose multiple description coding (MDC) as a way of introducing structured correlation among uncorrelated packets. By combining MDC and NC, we aim to ensure a reconstruction quality that improves gradually with the number of received network-coded packets. We consider two different approaches for generating descriptions. The first technique consists in generating multiple descriptions via a real-valued frame expansion applied at the source before quantization. Data recovery is then achieved via the solution of a mixed integerlinear problem. The second technique uses a correlating transform in some Galois field in order to generate descriptions, and decoding involves a simple Gaussian elimination. Such schemes are particularly interesting for multimedia contents delivery, such as video streaming, where quality increases with the number of received descriptions.Another application of such schemes would be multicasting or broadcasting data towards mobile terminals experiencing different channel conditions. The channel is modeled as a binary symmetric channel (BSC) and we study the effect on the decoding quality for both proposed schemes. Performance comparison with a traditional NC scheme is also provided.Concerning natural redundancy, a typical scenario would be a wireless sensor network, where geographically distributed sources capture spatially correlated measures. We propose a scheme that aims at exploiting this spatial redundancy, and provide an estimation of the transmitted measurement samples via the solution of an integer quadratic problem. The obtained reconstruction quality is compared with the one provided by a classical NC scheme.

This thesis treats the problem of source coding in situations where the encoded data is subject to errors. The typical scenario is a communication system, where source data such as speech or images should be transmitted from one point to another. A problem is that most communication systems introduce some sort of error in the transmission. A wireless communication link is prone to introduce individual bit errors, while in a packet based network, such as the Internet, packet losses are the main source of error. The traditional approach to this problem is to add error correcting codes on top of the encoded source data, or to employ some scheme for retransmission of lost or corrupted data. The source coding problem is then treated under the assumption that all data that is transmitted from the source encoder reaches the source decoder on the receiving end without any errors. This thesis takes another approach to the problem and treats source and channel coding jointly under the assumption that there is some knowledge about the channel that will be used for transmission. Such joint source--channel coding schemes have potential benefits over the traditional separated approach. More specifically, joint source--channel coding can typically achieve better performance using shorter codes than the separated approach. This is useful in scenarios with constraints on the delay of the system. Two different flavors of joint source--channel coding are treated in this thesis; multiple description coding and channel optimized vector quantization. Channel optimized vector quantization is a technique to directly incorporate knowledge about the channel into the source coder. This thesis contributes to the field by using channel optimized vector quantization in a couple of new scenarios. Multiple description coding is the concept of encoding a source using several different descriptions in order to provide robustness in systems with losses in the transmission. One contribution of this thesis is an improvement to an existing multiple description coding scheme and another contribution is to put multiple description coding in the context of channel optimized vector quantization. The thesis also presents a simple image coder which is used to evaluate some of the results on channel optimized vector quantization. / QC 20101108

Dans les réseaux traditionnels, la transmission de flux de données s'effectuaient par routage des paquets de la source vers le ou les destinataires. Le codage réseau (NC) permet aux nœuds intermédiaires du réseau d'effectuer des combinaisons linéaires des paquets de données qui arrivent à leurs liens entrants. Les opérations de codage ont lieu dans un corps de Galois de taille finie q. Aux destinataires, le décodage se fait par une élimination de Gauss des paquets codés-réseau reçus. Cependant, dans les réseaux sans fils, le codage réseau doit souvent faire face à des erreurs de transmission causées par le bruit, les effacements, et les interférences. Ceci est particulièrement problématique pour les applications temps réel, telle la transmission de contenus multimédia, où les contraintes en termes de délais d'acheminement peuvent aboutir à la réception d'un nombre insuffisant de paquets, et par conséquent à des difficultés à décoder les paquets transmis. Dans le meilleurs des cas, certains paquets arrivent à être décodés. Dans le pire des cas, aucun paquet ne peut être décodé.Dans cette thèse, nous proposons des schémas de codage conjoint source-réseau dont l'objectif est de fournir une reconstruction approximative de la source, dans des situations où un décodage parfait est impossible. L'idée consiste à exploiter la redondance de la source au niveau du décodeur afin d'estimer les paquets émis, même quand certains de ces paquets sont perdus après avoir subi un codage réseau. La redondance peut être soit naturelle, c'est-à-dire déjà existante, ou introduite de manière artificielle.Concernant la redondance artificielle, le codage à descriptions multiples (MDC) est choisi comme moyen d'introduire de la redondance structurée entre les paquets non corrélés. En combinant le codage à descriptions multiples et le codage réseau, nous cherchons à obtenir une qualité de reconstruction qui s'améliore progressivement avec le nombre de paquets codés-réseau reçus.Nous considérons deux approches différentes pour générer les descriptions. La première approche consiste à générer les descriptions par une expansion sur trame appliquée à la source avant la quantification. La reconstruction de données se fait par la résolution d'un problème d' optimisation quadratique mixte. La seconde technique utilise une matrice de transformée dans un corps de Galois donné, afin de générer les descriptions, et le décodage se fait par une simple éliminationde Gauss. Ces schémas sont particulièrement intéressants dans un contexte de transmission de contenus multimédia, comme le streaming vidéo, où la qualité s'améliore avec le nombre de descriptions reçues.Une seconde application de tels schémas consiste en la diffusion de données vers des terminaux mobiles à travers des canaux de transmission dont les conditions sont variables. Dans ce contexte, nous étudions la qualité de décodage obtenue pour chacun des deux schémas de codage proposés, et nous comparons les résultats obtenus avec ceux fournis par un schéma de codage réseau classique.En ce qui concerne la redondance naturelle, un scénario typique est celui d'un réseau de capteurs, où des sources géographiquement distribuées prélèvent des mesures spatialement corrélées. Nous proposons un schéma dont l'objectif est d'exploiter cette redondance spatiale afin de fournir une estimation des échantillons de mesures transmises par la résolution d'un problème d'optimisation quadratique à variables entières. La qualité de reconstruction est comparée à celle obtenue à travers un décodage réseau classique. / While network data transmission was traditionally accomplished via routing, network coding (NC) broke this rule by allowing network nodes to perform linear combinations of the upcoming data packets. Network operations are performed in a specific Galois field of fixed size q. Decoding only involves a Gaussian elimination with the received network-coded packets. However, in practical wireless environments, NC might be susceptible to transmission errors caused by noise, fading, or interference. This drawback is quite problematic for real-time applications, such as multimediacontent delivery, where timing constraints may lead to the reception of an insufficient number of packets and consequently to difficulties in decoding the transmitted sources. At best, some packets can be recovered, while in the worst case, the receiver is unable to recover any of the transmitted packets.In this thesis, we propose joint source-network coding and decoding schemes in the purpose of providing an approximate reconstruction of the source in situations where perfect decoding is not possible. The main motivation comes from the fact that source redundancy can be exploited at the decoder in order to estimate the transmitted packets, even when some of them are missing. The redundancy can be either natural, i.e, already existing, or artificial, i.e, externally introduced.Regarding artificial redundancy, we choose multiple description coding (MDC) as a way of introducing structured correlation among uncorrelated packets. By combining MDC and NC, we aim to ensure a reconstruction quality that improves gradually with the number of received network-coded packets. We consider two different approaches for generating descriptions. The first technique consists in generating multiple descriptions via a real-valued frame expansion applied at the source before quantization. Data recovery is then achieved via the solution of a mixed integerlinear problem. The second technique uses a correlating transform in some Galois field in order to generate descriptions, and decoding involves a simple Gaussian elimination. Such schemes are particularly interesting for multimedia contents delivery, such as video streaming, where quality increases with the number of received descriptions.Another application of such schemes would be multicasting or broadcasting data towards mobile terminals experiencing different channel conditions. The channel is modeled as a binary symmetric channel (BSC) and we study the effect on the decoding quality for both proposed schemes. Performance comparison with a traditional NC scheme is also provided.Concerning natural redundancy, a typical scenario would be a wireless sensor network, where geographically distributed sources capture spatially correlated measures. We propose a scheme that aims at exploiting this spatial redundancy, and provide an estimation of the transmitted measurement samples via the solution of an integer quadratic problem. The obtained reconstruction quality is compared with the one provided by a classical NC scheme.

With the era of wirelessly connected Internet of Things (IoT) devices on the horizon, eective data processing algorithms for IoT devices are of increasing importance. IoT devices often have limited power and computational resources, making data processing on the device unfeasible. Computational ooading, where the raw data is transmitted to a separate server, places a high load on the communication network, which in some cases may be prohibitively expensive. A split computing framework where some data pre-processing is done on the device, but the bulk of computations are done on a server at the network edge, provides a compromise between these limitations. Here, we employ a split computing framework in a semantic communication setting, where the semantic task is image classification. The system should fulfill three design requirements: low computational load on the IoT device, low load on the communication network, and good classification performance. We investigate the performance of two neural network structures: the first network is based on the VGG16 image classification network, and the second is the VGG16 network is augmented by separate encoder and decoder networks. The results are promising under both ideal and non-ideal channel conditions, where the first network gives good classification performance and low load on the communication network. The second network has low load on the IoT device, but surprisingly poor classification performance. Finally, we provide important insights into design choices and pitfalls, particularly reagrding network architecture and training, and hope that these results can aid future work in semantic communication systems. / I takt med att allt fler av våra system kopplas upp för kommunikation via internet, så kallad Internet of Things (IoT), får eektiva databehandlingsalgoritmer för dessa enheter av allt större betydelse. IoT-enheter har ofta begränsat minne, batteritid, och beräkningsresurser, vilket försvårar databehandling på enheten. Beräkningsavlastning, där rådata skickas till en separat server för behandling, kan leda till en hög belastning på kommunikationsnätverket, vilket i vissa fall är kostsamt. Att dela upp beräkningarna, där viss bearbetning av data görs på enheten men huvuddelen av beräkningarna görs på en server, är kompromiss mellan dessa två begränsningar. Här använder vi ett delat beräkningsramverk för semantisk kommunikation, där den semantiska uppgiften är bildklassificering. Systemet ska uppfylla tre designkrav: låg arbetsbelastning på IoT-enheten, låg belastning på kommunikationsnätverket och god klassificeringsprestanda. Vi undersöker två neurala nätverksstrukturer: den första är baserad på bildklassificeraren VGG16, och i den andra är VGG16-nätverket utökat med separata kodar- och avkodarnätverk. Resultaten är lovande under både ideala och icke-ideala förhållanden i kommunikationskanalen, där det första nätverket ger god klassificeringsprestanda och låg belastning på kommunikationsnätverket. Det andra nätverket har låg belastning på IoT-enheten, men överraskande dålig klassificeringsprestanda. Vi ger även viktiga insikter i designval och fallgropar, specifikt gällande nätverkens arkitektur och träning, och hoppas att dessa resultat kan gagna framtida arbete inom semantiska kommunikationssystem.