考慮人類視覺系統之三維轉換影像編碼

Abstract

在過去幾年當中，有許多有關人類視覺系統特徵之實驗結果被提出。有許 多研究將這些結果應用於影像編碼上，以改善視覺上的品質，在這方面， 大部份的研究都是用人類視覺系統的差異感測(contrast sensitivity)特 性。至今，空間域(spatial domain) 的差異感測函數( CSF or MTF)已 經被廣泛的應用，而且有效地改良編碼後的影像品質。採用差異感測函數 的典型編碼架構有：使用離散餘弦轉換(DCT)的靜態影像編碼，及利用運 動估計 (motion estimation) 和離散餘弦轉換的動態影像編碼。然而， 空間域的特性並不含括所有人類視覺系統的特徵。在時域 (temporal frequency)上的特性和其它的一些因素也扮演了很重要的角色。在這一份 研究報告中，我們考慮將一個結合空間和時間變化的差異感測函數加在以 三維離散餘弦轉換(3D DCT)為基礎的動態影像壓縮器內。量化器 (quantizer)的設計則考慮到率失真函數(rate-distortion function)和 差異感測函數對於不同信號的加重(weighting)特性。在給定的一個失真 上限下，我們可以調整差異感測函數來觀察所需的位元率(bit rate)和主 觀影像品質的互動關係。我們在編碼器中所用的只是一個簡化的模型，我 們主要的目的是想藉此探討三維差異感測函的視覺模型對主觀影像品質的 影響，因此我們才採用這樣的一個簡單模型。 更進一步的研究可以利用 這個基本模型加入其它的人類視覺特徵用於更複雜的編碼方法，利如以運 動估計為基礎的轉換編碼 (transform coding)。 Over the past few years, many empirical results on human visual system(HVS) behaviors have been reported. Many have also applied these results to image coding for improved perceptual quality. In this regard, most applied the contrast sensitivity characteristics of the HVS to their work. By now, the use of spatial-domain contrast sensitivity function (CSF, or MTF-the modulation transfer function) has been widespread and such use has indeed led to significantly improved coded picture quality. Typical coding frameworks in which the CSF is used include still picture coding employing the discrete-cosine transform (DCT) and motion-video coding employing motion estimation(ME) and DCT. Spatial-domain properties, however, do not characterize human visual perception fully. Temporal frequency and other factors also play a role. In this work, we consider the use of a combined spatial and temporal CSF (called three- dimensional CSF or 3D CSF) to the compression of motion video based on 3D DCT coding. The quantizer design is based on the rate-distortion function and the weighting property of the CSF. Given the distortion bound, we can adjust the CSF to see the relationship between the bit rate and the subjective quality. Since it is too complicated to exploit all human visual perception character- istics, the HVS model we use is a simplified one in many aspects. Our main aim is to explore the effect of a 3D CSF HVS on the subjective quality of coded images. Thus we have consi- dered a conceptually relative simple coding method. Further researches can utilize the concept of 3D CSF and other HVS char- acteristics in more complex coding methods, e.g., motion- estimation-based transform coding.