Sub-Optimal Context Modeling and Joint-Source-Channel-Code for JPEG Image Transmission with Unequal Error Protection

Jigisha N. Patel, Suprava Patnaik, and Manoj C. Joria




Context Modeling is a divide-and-conquer method based on separation of data into sub sequences by context, so that each subsequence can be approximated with a simple model while still providing a good overall precision. In this paper we have proposed context modeling of DCT based JPEG compressed image coefficient, suitable for joint source-channel coding. Proposed method relies on unequal error protection and achieves better perceptual quality as well as PSNR when transmitted using noisy channels. The conventional error control coding schemes uses Equal Error Protection (EEP), appending redundancy to all the information bits evenly. On the contrary Unequal Error Protection (UEP) aims at assigning protection relational to the significance of the information content of a pixel. Rate Compatible Punctured Convolution (RCPC) code is suitable for UEP application. UEP consists of mainly two parts: context modeling and compatible coding. In this work we have implemented two UEP approaches. In the first approach abbreviated as UEP_F, discrete cosine transform (DCT) coefficients are partitioning into DC coefficients and AC coefficients. In the second approach namely UEP_S, AC coefficients are further modeled, taking information saliency into account. Image sub-blocks from smoother regions are assigned to the high context model, whereas those with higher harmonics are tagged with low context model. The simulation is done for fixed transmission rate over an AWGN channel. The proposed method is called sub-optimal because no step has been taken to derive results for more number of models or précised classification which might have been the optimal. The result shows that proposed algorithm UEP_S improves the received image quality perceptually as well as in PSNR as compare to EEP and UEP_F for a given bit rate and noise conditions.

Important Links:

Go Back