基於最小可覺差之感知性視訊編碼

Abstract

基於最小可覺差之感知性視訊編碼可以移除感知上的多餘性並大幅增加編碼效率。然而大多現存之方法並不容易運用在位元率-誤差最佳化的架構中。在此論文中提出一基於最小可覺差模型之感知性視訊編碼並用其設計感知無失真視訊編碼系統以及感知失真視訊編碼系統。感知無失真視訊編碼系統是在感知無失真的前提下使用最少位元率來編碼，此方法藉由一自適應量化步階與動態量化位移選擇技術所達成。感知失真視訊編碼系統是藉由找出在編碼模式選擇與量化過程中位元率與感知品質的最佳權衡來達成。感知無失真視訊編碼系統可以在和H.264/AVC相同編碼條件下達到16%位元率減少，並透過主觀測試證明編碼後的視訊是感知無失真。感知失真視訊編碼系統可以在和H.264/AVC相同的感知品質下達到3-4%位元率減少。

Just-noticeable distortion (JND) model can be applied to remove the perceptual redundancies of video signal. However, most existing video coding systems use it for residual filtering or for replacing the conventional distortion metric. In this thesis, we present a JND-based approach and incorporate it into the rate-distortion optimization framework of video coding. Perceptually lossless video coding is achieved by minimizing bitrate subject to zero perceptual distortion. This is made possible by a quantization with adaptive step size and dynamic rounding offset. On the other hand, perceptually lossy coding is obtained by determining an optimal tradeoff between the bitrate and the perceptual quality distortion for mode decision and quantization. On the average, the proposed method as a perceptually lossless coding technique achieves 16% bitrate reduction compared to the H.264/AVC reference encoder under the same coding condition. A subjective test further confirms that the output video is perceptually lossless. As a perceptually lossy coding technique, the method achieves 3-4% bitrate reduction over the H.264/AVC reference encoder at the same visual quality level.