本篇論文提出一個修補影片的系統,其主要的偵測對象是短暫性出現的雜訊。我們利用以區塊為基礎的移動估計,使用SAD(Sum of Absolute Difference)公式計算移動向量,進而偵測出雜訊。接著再將偵測到的雜訊過濾,最後再利用temporal及spatial的資訊雙重修補。Spatial的修補方法是利用多重解析度影像修補(multi- resolution image inpainting)。 本研究的結果,可以透過我們的偵測與過濾的演算法,成功地將毀損的區域偵測並修補;大部分的雜訊都可以成功的被偵測到,偵測率約為90%;而其修補後的結果也非常平滑、影片瀏覽順暢,視覺效果與原本影片非常相近,實作結果很好。
Video inpainting uses spatial-temporal information to repair defects such as spikes and lines on aged films. We propose a series of new algorithms based on adjustable thresholds to repair different varieties of aged films. The main contribution is an automatic spike and dirt detection mechanism. We prove that if appropriate threshold is once decided by the author, almost all damages in an aged video clip can be detected. In addition, the repairing procedure first estimates temporal information and obtain replacement blocks among several frames. Spatial information is then used to repair damages that can not be fixed by temporal information due to fast motion. The results are visually pleasant with most defects removed.