Multi-reference neighborhood search for vector quantization by self-organized featured map

摘要

Reference neighbor search (RNS) is a new hybrid coding technique using the DCT and vector quantization (VQ) for fast searching, while maintaining near-optimal visual quality. However, the performance is greatly affected by the selection of reference point. We employ a self-organized feature map to generate a topological preserving codebook and adaptively predicts the reference by a simple scheme. The predicted reference is close to the input, together with searching across multiple queues. The searching cost was significantly reduced while maintaining optimal visual quality.