This work is devoted to practical joint source channel coding. Although the proposed approach has more general scope, for the sake of clarity we focus on a specific application example, namely, the transmission of digital images over noisy binary-input output-symmetric channels. The basic building blocks of most state-of the art source coders are: 1) a linear transformation; 2) scalar quantization of the transform coefficients; 3) probability modeling of the sequence of quantization indices; 4) an entropy coding stage. We identify the weakness of the conventional separated source-channel coding approach in the catastrophic behavior of the entropy coding stage. Hence, we replace this stage with linear coding, that maps directly the sequence of redundant quantizer output symbols into a channel codeword. We show that this approach does not entail any loss of optimality in the asymptotic regime of large block length. However, in the practical regime of finite block length and low decoding complexity our approach yields very significant improvements. Furthermore, our scheme allows to retain the transform, quantization and probability modeling of current state-of the art source coders, that are carefully matched to the features of specific classes of sources. In our working example, we make use of ``bit-planes'' and ``contexts'' model defined by the JPEG2000 standard and we re-interpret the underlying probability model as a sequence of conditionally Markov sources. The Markov structure allows to derive a simple successive coding and decoding scheme, where the latter is based on iterative Belief Propagation. We provide a construction example of the proposed scheme based on punctured Turbo Codes and we demonstrate the gain over a conventional separated scheme by running extensive numerical experiments on test images.