Bitwise Structured Prediction Model for Lossless Image Coding

In this paper, we propose the bitwise structured prediction model for the loss less image coding. The prediction problem is handled by decomposing into a series of bitwise prediction problems, where max margin estimation is made for each bit. Furthermore, the structured prediction is proposed for combining such bitwise problems with constraints on their interrelationship and loss function for loss-augmented inference. The proposed methods tend to utilize the inherent dependencies existing in the bitplanes among the neighboring pixels and suppress the fluctuation led by the bitwise decomposition. When building the max margin Markov network for training, the upper bound for the prediction errors is shown to be asymptotically equivalent to the results obtained over the training set. Experimental results also show that the proposed method is superior in coding performance for regular oscillatory patterns.