Mismatch Of Inverse Discrete Cosine Transform And Its Fixed-point Implementation

In conventional video coding standards based on hybrid coding technology, transform coding plays an essential role. Among various orthogonal transforms, Karhunen-Loeve transform (KLT) exhibits the best compression ability while discrete cosine transform (DCT) has less complexity but similar performance. Due to many existing fast algorithms, DCT is widely used in many video coding standards.The fundamental of DCT is introduced and some fast algorithms are analyzed at the beginning of this paper, followed by introduction of recent standardization activity of fixed-point8x8DCT/IDCT (ISO/IEC23002) in MPEG video group. We give some examples of currently existing fixed-point IDCT implementations with detailed analyse of their characteristics and performances.Meanwhile, drift problem and some potentioal facts to aggravate it are discussed in this paper. Further, two types of IDCT mismatch artifacts under certain coding conditions are presented. We conclude the causes of those severe drift problems as single pixel error accumulation and error enlargement and propagation with sub-pixel interpolation. Moreover, we proposed three ways to avoid or counteract those artifacts and improve the quality of reconstructed pictures, where IDCT mismatch cannot be eliminated.To avoid mismatch fundamentally, the only way is to set one uniform, bit-exact IDCT implementation on both encoder and decoder, which is also the aim of ISO/IEC23002-2. The author collaborated with other professors to propose a globalized optimization method to choose the best appropriate fixed-point IDCT coefficients. Three fixed-point IDCT implementations with different accuracies and complexities are provided as examples of that method. Very high accuracy is achieved at the expense of low complexity. This method has been adopted by ISO/IEC23002-2.