適用於即時行動視訊之頻寬導向移動估測研究

Abstract

隨著對行動裝置視訊服務的上升趨勢，如何將這些裝置的功率消耗減到最小並延長使用的壽命成為現今的主要設計問題。其中，視訊壓縮為主要的功率消耗來源，特別是移動估測。為了最好的位元失真率性能計算及參考資料的存取，移動估測是計算密集及高頻寬需求，但由於行動裝置易變的操作情況及可獲得的資源，這些在滿足有限制功率的行動裝置是很困難的。 這篇博士論文提出可調整資料頻寬的移動估測觀念來應付這些問題，同時使移動估測資源的使用適應可獲得的資源及操作情況，且儘可能的最佳化或維持品質。根據可獲得的資源及操作情況，這篇博士論文藉由不同的位元失真率與資料頻寬模型提出三種移動估測設計。 第一個設計為有警覺資料頻寬的移動估測設計，藉由合併資料頻寬的條件到傳統的位元失真率分析，使它的頻寬使用適應到可獲得的資源。此方案動態的預測與分配資料頻寬，同時根據位元失真率增益與資料頻寬效率來最大化位元失真率性能。與H.264參考軟體的比較，模擬結果顯示針對低移動量CIF尺寸的視訊，此方案可節省70%的頻寬同時維持相同的位元失真率性能。結果顯示針對高移動量CIF尺寸的視訊，此方案在相同的頻寬使用下，可以善加使用可獲得的頻寬並節省高達13%的頻寬與增加高達0.1分貝的峰值信噪比。

第二個設計為有限制資料頻寬的移動估測設計，此設計在同時嚴格的頻寬條件與相同的品質性能條件下，更進一步的縮小資料頻寬的使用。此方案針對低移動量CIF與D1尺寸的視訊與高移動量CIF與D1尺寸的視訊，可以分別節省高達58%到66% 與37%到52%的資料頻寬，且同時維持相似的品質。這個設計以台積電的0.13微米互補式金屬氧化物半導體製程，花費了122K的邏輯閘數目與21.5K位元組的SRAM，同時針對每秒30幀的D1解析度，可以節省60%的外部記憶體存取功率。 最後一個設計為有限制資料頻寬的移動估測設計，藉由公式化位元失真率成本與可獲得的資料頻寬為凸集最佳化架構來最佳化資源使用，所提的方案藉由前一幅相同位置區塊的移動力估測來決定最佳的搜尋範圍。實現的結果顯示所提位元失真率成本相對搜尋範圍的模型可以精確的預測位元失真率成本趨勢與較低的外部記憶體存取，與參考軟體JM相比，在相似的位元失真率性能下，可節省平均51%的頻寬。 總結，我們可以採用上面所提的可調整資料頻寬的移動估測來緩和行動視訊裝置的設計負擔與達到低功率與即時的運用。

With the rising trend for video services on mobile device, how to minimize the power consumption and prolong the operational lifetime of these devices is the primary design issue nowadays. In which, the majority of power consumption is due to video compression, especially on motion estimation (ME). ME is computational intensive for best rate-distortion performance calculation and bandwidth-demanding for reference data accesses, which is hard to meet on a power-limited mobile device due to its volatile operating conditions and available resources. To address the problems, this dissertation proposes the data bandwidth-scalable ME to adapt the resource usage to the available resources and operating conditions while maximize or maintain the quality as good as possible. According to the available resources and operating conditions, we propose three ME designs in this dissertation by different models of rate-distortion and data bandwidth. The first design is a bandwidth-aware ME that can adapt its bandwidth usage to available resources by incorporating the data bandwidth into conventional rate-distortion analysis. It dynamically predicts and allocates the data bandwidth while maximizing the R-D performance according to its R-D gain and data bandwidth efficiency. The simulation result shows 70% BW savings while keeping equivalent R-D performance compared with H.264 reference software for low-motion CIF-sized video. For high-motion sequences, the result shows this scheme can better use the available BW to save an average bit rate of up to 13% with up to 0.1-dB PSNR increase for similar BW usage. The second design is a bandwidth-constrained ME that further minimizes the data bandwidth usage under strict bandwidth constraint with linear bandwidth-rate-distortion model. This scheme can save data bandwidth up to 58% and 66% for low-motion CIF- and D1-size videos, and up to 37% and 52% for high-motion CIF- and D1-size videos with similar quality. This design costs 122 K gate counts with TSMC 0.13-μm CMOS technology and 21.5 KB SRAM, and can save 60% external memory access power for D1 resolution at 30 frames/s. The final design on bandwidth-constrained ME optimizes the resource usage by formulating the R-D cost and available data bandwidth into a convex optimization framework. The proposed scheme determined the optimum search range by motion activity estimation from the co-located MB in the previous frame. Simulation results demonstrate that the proposed RDcost-SR model can precisely predict the R-D cost trend and lower external memory access with on average 51% bandwidth saving at the similar R-D performance compared to reference software JM. In summary, with above bandwidth-scalable ME scheme, we can relieve the design burden of mobile video devices and achieve lower-power and real-time applications.