解交錯演算法使用平滑條件適應性選擇空間與時間的插補

Abstract

本論文中，我們呈現一個動態選擇空間插補與動態補償的混和式 的解交錯演算法。在空間插補的演算法中，我們參考了影像補償的觀 念，等光線，估計出大約的平滑方向且在稍後的步驟中正確的找出與 粗估的最相近的方向。時間插補方法中，動態補償是主要的時間插補 的結果。在動作估計中，我們增加一個平滑限制的因子使得由動作估 計呈現出動作向量的連續性而且也會產生出較好的動態補償結果。 為了避免因為錯誤動作向量所產生的嚴重區塊失真，利用動作向 量能量與自相關函數去偵測哪一個向量是否是錯誤的。像素值應該與 其相鄰的像素有關連性是眾所周知的。把像素基礎的亮度的關聯性轉 換成以區塊基礎的動作向量，我們引用了兩種不連續的類型去估計動 作向量的能量：區域與上下文的不連續。一般而言，大的動作向量能 量值越大表示此向量越不可靠。 在我們的方法中，大量的平滑邊與區塊失真都被正確的找出。選 擇插補的方法依靠動作資訊而不採用空間特性且最後的輸出是空間 與時間插補的二分選擇。實驗的PSNR 數據比現有新穎的演算法至少 有0.5db 的增加。

In this thesis, we present a hybrid de-interlacing algorithm using spatial interpolation and motion compensation. The spatial interpolation refers to the concept of image inpainting, isophote, to estimate smooth direction approximately and then corrected by smoothness factor further. In temporal interpolation, motion compensation (MC) is the major temporal result. In the motion estimation (ME), we add a factor of smoothness constraint to present the continuity of motion vector (MV) and it would produce better motion compensated result. In order to prevent the signi…cant block e□ect caused from erroneous MV, using the energy of MV and autocorrelation function to detect which MV(s) is erroneous. It is well known that pixel value has correlation with its nighborhoods. To translate the correlation from pixel-based intensity value to block-based MV length, we introduce the two types of discontinuity on MV to evaluate the motion vector energy (MVE) of MV: local and contextual discontinuity. Generally, the large measure score of MVE means that the MV is unreliable, vice versa. The significant smooth edges in spatial method and block artifacts in temporal results are recognized correctly. The manner of choosing interpolation depends on the motion information rather than spatial feature and the …nal output is a binary decision between spatial and temporal interpolation. The simulated peak signal-to-noise ratio (PSNR) of proposed algorithm have at least 0.5dB better than other novel de-interlacing algorithms.