量化失真模型對影像壓縮結果的訊號-雜訊比的影響

Abstract

在TMN8的速率控制機制中，是要藉由建立速率模型跟失真模型，來選擇一組量化參數，以使得在壓縮速率的限制下，達到失真最小的結果。因此，若使用一個較準確的失真模型，則可使得壓縮出來的影像失真較小。在TMN8中原本使用的失真模型，是Uniform失真模型，但其描述的失真分布跟實際的失真分布有所落差。所以，我們在此篇論文中，在假設DCT係數是Laplacian分布的情形下，推導了Laplacian失真模型。我們可看出，Laplacian失真模型的確較接近實際失真分布。而且，在TMN8中使用Laplacian失真模型，的確可以獲得較小的MSE失真，因而使得影像壓縮的訊雜比較好。

In TMN8 rate control structure, both the rate model and the distortion model are used to calculate the quantization parameters such that the MSE distortion is minimized subject to the constraint that the resultant rate meet the target bit rate. However, the distortion model in TMN8 is based on the assumption that the DCT value follows the uniform distribution. This assumption is not consistent with the actual distortion, thus with TMN8 the actual distortion is not minimized. In this thesis Laplacian distortion model is proposed based on the fact that the DCT value follows the Laplacian distribution. We have first shown that the Laplacian distortion model is actually more close to the actual distortion. Thus with the Laplacian distortion model in TMN8 rate control, the resutant MSE distortion is smaller than that with the uniform distortion model under the same target bit rate.