本研究計畫提出基於內容適應性與編碼效果預測之H.264/AVC視訊壓縮機制。透過由視訊內容中擷取之時間域與空間域特徵，以及本研究所提出的位元率-量化參數與品質-量化參數模型，我們可於視訊編碼過程中即決定適當的量化參數，有效率地執行H.264之畫面內編碼與畫面間編碼。量化參數所產生的畫面位元率與畫質能夠於壓縮前予以較準確地預測，有助於視訊位元率的控制。此外，本研究更將所提出的壓縮機制應用於兩種位元率控制類別，即穩定畫質視訊編碼與穩定位元率視訊編碼。在穩定畫質視訊編碼中，我們藉由視訊片段邊界偵測，取出具有同樣特性的視訊片段，並以穩定畫質為目標壓縮該視訊片段。穩定畫質編碼適用於監控畫面，透過適當的編碼參數調整，我們可讓每張監控畫面具有相同的品質，以確保日後該畫面用於線索搜查時的實用性。此外，經由位元率-量化參數模型，我們亦可確保該片段經壓縮後所產生的位元率適當地符合監控視訊傳輸的頻寬要求。換句話說，本研究所提出之視訊壓縮模式將與相關應用結合，以彰顯研究的重要性。在穩定位元率視訊編碼中，我們希望經由適當的位元率分析給定該張畫面之資料量限制後，以適當的量化參數進行編碼，讓壓縮視訊的位元率符合目標需求。同樣的，藉由所提出的模型，我們可給予畫面適當的目標位元率，避免不同畫面間的畫質激烈改變。我們更提出基於內容分析方式所產生之有興趣區間編碼。我們建構關注模型以於畫面中擷取有興趣區間，並於該區間與背景分別給定適當的量化參數，讓有興趣區間的畫質受到更佳的保障。透過同一張畫面可能使用多個量化參數，我們可讓該畫面的目標位元率更精確地被達成。我們亦將此穩定位元率壓縮機制實作於監控視訊，讓監控目標具有較佳畫質，並且讓我們能夠更有彈性地調整量化參數已達成更準確的穩定位元率編碼，符合監控視訊應用之需求。我們的實作將與常見的位元率控制機制比較，以彰顯我們所提出之方法在畫質與位元率上的優勢。This research presents a content-adaptive performance-aware rate control scheme for H.264/AVC. Via the spatial and temporal feature extraction from the video content, the Rate-Quantization (R-D) and Distortion-Quantization (D-Q) models will be formed so that the suitable Quantization Parameters (QP) can be determined in both the intra- and inter-frames of H.264/AVC during the encoding processes. The term “Performance-aware” means that the resulting bit-rate and quality distortion associated with QP values can be predicted pretty accurately to facilitate the rate and quality control. The proposed content-based rate control models will be applied in both the cases of Variable-Bit-Rate (VBR) and Constant-Bit-Rate (CBR) controls. In VBR, our scheme first employs the shot boundary detection to determine the video shots so that the constant quality video coding can be applied in the shots. The target application of constant quality video coding is video surveillance. Through the appropriate setting of QP values, we can ensure that each video frame have reasonable quality such that the recorded video frames will become more useful if they will be examined to find the evidence of a certain event afterwards. By using our R-Q model, we can also ensure that the resulting bit-rate of the shot should still fulfill the bandwidth requirement of video surveillance. In other words, our rate control models will be combined with the target application to highlight the contribution of the proposed scheme. In CBR, after exploiting the content analysis to determine the suitable target bit-rate to make the video quality more consistent, we can use our proposed model to determine the QP values and to generate the bit-rate close to the objective. In this research, we will propose a content-based Region of Interest (ROI) video coding, in which an attention model is constructed for extracting the ROI of an input frame. By assigning different QP values to the ROI and background, we can ensure that the quality of ROI be well preserved and the resulting bit-rate should be set more accurately. Again, we will apply our scheme to the scenario of video surveillance. Our scheme can not only enable the target of surveillance to be processed more carefully but also have more flexibility of adjusting QP values to achieve the constant bit-rate coding. We will compare our scheme with the existing rate-control mechanisms to demonstrate the advantages and feasibility of our method, both in the frame quality and bit-rate. Furthermore, the requirements of the target applications can also be fulfilled. 研究期間：10008 ~ 10107
