兩種處理壓縮視訊內容查詢的有效方法: 過濾和預測

Abstract

在本篇論文中，我們提出過濾(filtering)和預測(prediction)兩種處理壓縮視訊內容查詢的有效方法。其中，過濾的方法是擷取視訊資料在壓縮領域下的離散餘弦函數轉換(Discrete Cosine Transform, DCT)係數，進行影像的相似比對工作，同時採用多層解析度(multi-resolution)的比對方式以達到過濾的功效。預測的方法則是利用壓縮視訊資料中的區塊(macrobalock)型態和運動向量(motion vector)等時間特徵，進行影像相似程度的預測，以減少計算DCT係數所需的時間，加速壓縮視訊內容查詢。在論文中，我們利用過濾和預測這兩種方法進行壓縮視訊資料的場景偵測(scene detection)、整張影像(full image)的相似比對以及部分範例的查詢工作(query by partial example)。過濾和預測此兩種方法，可在不解壓的前提下，有效地減少視訊內容相似比對所需的時間。視訊資料查詢功能的部分，除了提供在壓縮條件下整張影像查詢的功能，本篇論文亦提出採取區塊相似比對法則，來提供部分範例查詢的功能，以有效地擷取相似的視訊物件。

In this paper we propose two efficient approaches namely filtering and prediction for content-based retrieval of compressed video data. For filtering, we retrieval the DCT coefficients of video data in compressed domain to proceed similarity comparison. Concurrently, a multi-resolution comparison method is adopted to proceed filtering. As for prediction, we use macroblocks and motion vector, which are temporal features of compressed video data, to predict similarity of video data such as to reduce the amount of time for computing and speed up content-based retrieval of compressed video data. The proposed approaches can efficiently reduce the time needed for content-based retrieval without using decompression. For real applications, we successfully applied filtering and prediction to scene detection, query by full image, and query by partial example of compressed video data.