Entropy Coding for Low-Bit-Rate Visual Telecommunications

Several compression techniques need to be integrated for the achievement of effective low-bit-rate coding of moving images. Image entropy codes are used in conjunction with either predictive or transform coding methods. In this paper, we investigate the possible advantages of using arithmetric codes for image entropy coding. A theory of source modeling is established based on the concept of source parsing and conditioning trees. The key information-theoretic properties of conditioning trees are discussed along with algorithms for the construction of optimal and suboptimal trees. The theory and algorithms are then applied to evaluating the performance of entropy coding for the discrete cosine transform coefficients of digital images from the "Walter Cronkite" video sequence. The performance of arithmetic codes is compared to that of a traditional combination of run length and Huffman codes. The results indicate that binary arithmetic codes outperform run length codes by a factor of 55 percent for low-rate coding of the zero-valued coefficients. Hexadecimal arithmetic codes provide a coding rate improvement as high as 28 percent over truncated Huffman codes for the nonzero coefficients. The complexity of these arithmetic codes is suitable for practical implementation.