動態自動對焦演算法

Abstract

自動對焦，是數位影像加強技術中重要的議題之一。近年來數位相機/攝影機的普及，隨著科技的進步，影像品質逐漸受到更多的重視。影像加強是影像處理技術的核心，影像加強的目的是追求真實，修正影像與人眼感官相符。自動對焦和自動曝光、自動白平衡，並稱3A演算法，是目前影像加強的主要演算法。靜態影像的自動對焦演算法已經有很多的研究成果，唯動態影像的自動對焦演算法尚沒有顯著的討論。本篇論文改良整合現有的自動對焦技術，提出一個動態自動對焦的演算法。

動態自動對焦與靜態自動對焦的主要差異，在於搜尋演算法沒有所謂起始和終止，必須不斷地遞迴搜尋。論文先分析數位攝影機的基本架構，並討論其回饋系統的特性。討論自動對焦的結果，會受到系統本身哪些的限制。我們針對畫面轉換、物體晃動、焦距變動幾個現象，改良現有的爬山搜尋演算法，維持自動對焦系統的效能和穩定性。我們以有限狀態機器的形式，藉由設定專門的轉換參數，控制搜尋演算法的狀態轉換，來實現動態自動對焦演算法。我們利用亮度指標來判斷換面轉換的發生，利用門檻加強的搜尋演算法，處理物體晃動的問題。並提出一個線性的預測模型，來改善焦距變動的對焦效率。最後我們以軟體模擬的形式，來驗證我們的自動對焦演算法能對動態影像運作。

Auto-Focus is one of the most significant research issues in the recent digital image enhancement technology. Its importance increases due mainly to the widespread use of digital video/imaging devices in this new millennium. The stationary image auto-focusing system is a long-standing research topic. In this thesis, we modify and improve the conventional auto-focusing algorithms and integrate them into a dynamic auto-focusing algorithm. The major difference between a still image and a video auto-focusing system is the search algorithm. Still image search algorithms often have specific start and end, but the video application urges a continuous searching routine. First, we study the digital camera system and the feedback control theory. Through these studies, we understand how the system structure limits the auto-focusing result and the principles of designing a good auto-focusing system, which is a special type of control system. Since the search algorithm is critical, we improve the climbing search algorithm for particularly the dynamic environments such as scene change, local object motion, and zoom tracking. This proposed search algorithm working on the video is the core of our auto-focusing algorithm. Then, we develop our search algorithm using a finite state machine structure. We design specific transition conditions and state transition table to match the requirements of dynamic auto-focusing applications. A luminance-based metric helps to detect the scene change. We adopt a threshold climbing search algorithm to solve the local motion problem. And the zoom tracking processing is accelerated with the assistance of a well-designed linear prediction model. Finally, we show that this algorithm is reliable and efficient by a series of software simulations.