Development of an Entropy Code for Component Based Image Compression

A major problem in image compression is the coding of noise and noisy texture in an image. In a classic image compression approach, a transparent representation of these components is only possible with a very high bit rate. However, a pixel exact representation is not necessary in order to achieve visually perceived transparency because the human viewer is not able to distinguish two noise signals with the same statistical properties. Ballé [1] proposes a system that allows for a reconstruction of noise without pixel wise exactness. At first, the image is decomposed into a structural part and another part that contains noise and noise like texture. The structure part is encoded using conventional techniques (e.g. JPEG 2000) where the noise component is modeled by an extended autoregressive (ARX) model. At the receiver side, we then use the ARX model coefficients to reconstruct a noise component with similar characteristics. In this paper, we propose and evaluate different coding schemes in order to design a fast and efficient entropy code for these ARX model coefficients. We also show a first comparison to a conventional image compression technique.