運用適應性分鏡關鍵特徵集合於多影片查詢

A Multi-Video Retrieval Using Adaptive Key Feature Set of Shot

曹盛魁(Sheng-Kuei Tsau)

指導教授 ： 林春宏

國立臺中科技大學/資訊與流通學院/資訊工程系碩士班/碩士(2012年)

https://doi.org/10.6826/NUTC.2012.00072

摘要

本文提出以關鍵特徵集合來描述影片分鏡內容，以此特徵集合透過特徵的距離量測方式，計算查詢影像(片)與影片資料庫間特徵距離，回傳相似結果，達成多影片查詢此一目標。
本研究針對網路上互動率高的影片對其特性分析，透過K-mean演算法及LBP之梯度直方圖特徵分別轉換為色彩及紋理特徵，萃取出可代表視訊頁(frame)的特徵直方圖，以此特徵計算連續視訊頁間距離，接著進行分鏡邊界偵測(shot boundary detection)，找出影片中分鏡轉換的邊界切割出不同分鏡。影片切割成具代表性的分鏡後，以影片分析概念將影片分鏡做過濾及分群動作，使各分鏡各具代表性，再將具影片代表性的分鏡建立關鍵特徵集合(key feature set)以作為影片描述。透過欲查詢影片及影片資料庫間的關鍵特徵集合，做距離計算即可找出相似的影片，以此體現多影片查詢(multi-video retrieval)此一目標，可提供使用者不同思維的查詢方式。
本研究所提出的影片查詢方法不同於現有影片查詢之處在於，目前能見度較高的線上視訊分享網站如YouTube、Yahoo video、VEVO等都是以文字為基礎查詢，本文查詢方式則是以色彩及紋理特徵所建立的關鍵特徵集合為基礎。在比對結果上不同於以文字來描述影片，而是以影像中的色彩及紋理內容為查詢，因此能避免文字上的理解方式不同而導致查詢的結果雜亂，對使用者來說以同樣的影像或影片去查詢自己感興趣的內容更為直覺。
在實驗結果中，將分別就分鏡邊界偵測上，以偵測出正確的分鏡轉換為目標，並與歷來研究者以數據做比較，以及影片分析對影片做分群，將分鏡內容中異同做分別，使之後建立的關鍵特徵集合更具代表性。最後影片查詢分別就影像查詢影片及影片查詢影片兩種方式回傳查詢目標相似的結果，可由實驗結果觀察得知，僅以影片片段即可查詢出相似特徵的影片。

關鍵字

分鏡邊界偵測 ； 關鍵特徵集合 ； K-mean演算法 ； 梯度二值化圖樣特徵 ； 影片查詢

並列摘要

In our thesis, we present different key feature sets to describe the content of every shot in video. Calculate the distance between query image (or query video) and video database by distance measure of feature. Return the most similar results to build a multi-video retrieval function.
In our study, we analyze the digital videos on the Internet. By K-mean algorithm and the gradient histogram of LBP, we respectively transferred them into color and texture features and extracted histogram of each frame. According to the features, the distance between continuous frames can be calculated and the shot boundary can be detected to identify the shot boundaries of the video. Thus, different shot can be created. After video were cut into representative shots, filter and cluster the shots by video analysis function, so that each shot can be more representative. According to the correct shots, key feature set can be built to describe the video. Finally we can find out the similar results by calculating the distance between query video and video database. After the steps above, the goal of multi-video retrieval can be reached and the query result can provide the user with a different way of video retrieval.
Where the video retrieval method proposed in this study differs from existing video retrieval is that the current high visibility online video sharing web sites such as YouTube, Yahoo video, VEVO and others are all text-based queries. The query function of our study is based on the color and texture features of the key feature set. These retrieval results differ from using words to describe the video. Instead, it searches with color and content of texture in video. Thus, chaos of the result caused by different text reading method can be avoided. The users can be more intuitive to query the video interested by using the same image or video.
In the experimental results, we will focus on detecting the correct shot transferring based on shot boundary respectively. By shot filter and shot cluster function in video analysis, we identified the difference of shot content, so that the key feature set can be more representative. Finally experimental result in video retrieval, it can be divided into two ways - the image query videos and the video query videos to return similar result of the query goals. From the results we know similar shot of the query target can be found simply by sections of video.

並列關鍵字

shot boundary detection ； key feature set ； K-mean algorithms ； gradient histogram of local binary pattern ； video retrieval

參考文獻

[8] D. Scharstein and R. Szeliski, "A taxonomy and evaluation of dense two-frame stereo correspondence algorithms," International journal of computer vision, vol. 47, pp. 7-42, 2002.

[9] A. Albiol, M. J. Silla, and J. M. Mossi, "Video analysis using corner motion statistics," Performance Evaluation of Tracking and Surveillance, pp. 31-38, 2009.

[10] M. Mrak, J. Calic, and A. Kondoz, "Fast analysis of scalable video for adaptive browsing interfaces," Computer Vision and Image Understanding, vol. 113, pp. 425-434, 2009.

[11] W. Chen and Y. J. Zhang, "Parametric model for video content analysis," Pattern Recognition Letters, vol. 29, pp. 181-191, 2008.

[13] G. Toderici, H. Aradhye, M. Pasca, L. Sbaiz, and J. Yagnik, "Finding meaning on youtube: Tag recommendation and category discovery," Computer Vision and Pattern Recognition, pp. 3447-3454, 2010.