視訊編碼之景緻分析技術

Abstract

有限位元率(bit rate)達到較好影像品質，需要同時使用主觀及客觀的壓 縮技術。本論文中、我們提出景緻分析的方法，進行主觀的壓縮，它使用 最少位元率在整張影像上產生視覺相同的失真，所以也改善壓縮影像的品 質。方法是在區塊離散餘弦轉換域(Block DCT domain)探討人類視覺系統 的三個特性：頻率感度 (frequency sensitivity)、背景感度(contrast sensitivity)及遮罩效應(masking effect)。在國際視訊標準中(如 H.261 及 MPEG1)，通常採用區塊離散餘弦轉換編碼技術，所以我們建議 及設計在此域中量測人類視覺特性的實驗。這些特性中、遮罩效應是決定 影像區塊感受界限(Perceptual Threshold)的主要因素，我們先將影像分 成四類，每類具有相同隱藏誤差的能力，因此使決定感受界限的複雜問題 簡單化。每類的區塊活性函數(Block Activity Function)定義為所有交 流係數(AC coefficients)的絕對值和，可作為分成數個細類的依據，而 每個細類具有相似的感受界限，論文中也包含這方面的相關實驗。上面的 結果可作為影像品質及位元率的控制。最後、我們使用已知的主觀失真量 測(Subjective Distortion Measure) 函數來衡量所提出的方法。 A better quality picture with a constrained bit rate can be achieved by using both the objective and the subjective compression techniques together. In this thesis, we propose a scene analysis method to implement subjective compression which leads to quality improvement on compressed images, because it would produce equally visible coding distortion across the entire picture using the least amount of bits. This method exploits three HVS properties in Block Discrete Cosine Transform (BDCT) domain, namely frequency sensitivities, contrast sensitivities, and masking effects. We suggest and conduct several experiments for measuring the HVS properties in BDCT domain. BDCT coding scheme has been adopted by several video international standards, for example, H.261 and MPEG1. Among these properties, spatial masking is the key factor in deciding the perceptual thresholds of picture in local regions. We develop a perceptual-based picture classification algorithm and classify picture regions into four categories. Each category has similar capacity in hiding errors, and thus making the complicated problem in determining perceptual thresholds more easily. The Block Activity Function (BAF) of each category, sum of the absolute values of AC coefficients, can be used to further divide a category into subcategories. As a result, each subcategory now has a quite uniform visibility threshold. Some experimental results are included in this thesis, which can be used for picture quality and bit rate control in coding. Finally, we use a subjective distortion measure to evaluate our proposed approach.