頻寬-位元率-失真最佳化之移動估測

Abstract

移動估測在H.264視訊編碼的過程中，具有龐大的運算量和記憶體需求，然而傳統的移動估測只考慮位元率和失真，未將記憶體頻寬納入考量，因此，在頻寬受限的情形下，位元率和失真並未做到最佳化。為了解決以上因素，在此本論文提出一個頻寬-位元率-失真最佳化的移動估測演算法，首先，我們提出一個頻寬-位元率-失真最佳化的模型，藉由設立一個合理的搜尋範圍，在允許的頻寬下提升位元率和失真的最大效能。其次，我們提出兩個方法來進一步提高前者模型的效能，一種方法為隨內容感測的跳躍預測演算法，藉由跳躍預測所節省的頻寬來提升其餘複雜畫面的編碼品質；另一方法為搜尋範圍邊界預測演算法，藉由設立一個適宜的搜尋範圍邊界而對搜尋範圍作進一步的修正。相較於之前的研究，對於靜態畫面影像，我們的設計在相同位元率與失真，以及平均搜尋範圍為16的情形下，頻寬可節省70%，若加上跳躍預測演算法，則頻寬可節省84%；對於動態畫面影像，我們的設計在低頻寬的環境下，位元率可降低13%，同時峰值信噪比也可提升0.1dB。總結我們的設計對於不同的頻寬環境變化下，不僅維持了效能，甚至更進一步提升了效能，這顯現出我們的設計可適用於改進移動估測的處理。

Motion estimation (ME) processing is the most computational and memory intensive component in H.264 encoder. However, traditional ME algorithms focus on rate and distortion performance and thus do not take memory bandwidth into consideration. Therefore, the rate and distortion performance are not optimized under bandwidth constraint. In this thesis, we propose bandwidth-rate-distortion (B-R-D) optimized ME algorithm to solve the issue mentioned above. First, we mainly propose a B-R-D optimized modeling method to determine an appropriate search range (SR) for maximizing rate distortion efficiency while can dynamically meet the available bandwidth. Then, we propose two methods, skip mode detection with content-aware scheme and SR boundary prediction method, to enhance the performance of B-R-D optimized modeling method. The skip mode detection with content-aware scheme is presented to save the most memory bandwidth and thus gives other complex MBs more bandwidth for better quality, and the SR boundary prediction method is presented to determine a feasible SR boundary for SR refinement. Compared with reference software [3], when coding in low motion sequence, the simulation result shows the proposed BRD design could improve the bandwidth saving up to 70% with almost the same performance at bit rate and PSNR under average search range size 16, and up to 84% with negligible PSNR degradation with skip design added; while coding in high motion sequence, the simulation result shows our design could save average bit rate up to 13% and at the same time increase average PSNR up to 0.1dB under low bandwidth constraint. In summary, our design could achieve the same and sometimes even better performance under various bandwidth constraints and thus it is suitable for improving ME process.