【摘要】： 隨著互聯(lián)網(wǎng)技術的迅速發(fā)展,網(wǎng)上的多媒體信息也越來越多,特別是近兩年來,數(shù)碼動畫繼音樂和圖片之后異軍突起,成為又一種互聯(lián)網(wǎng)上用來傳播信息的常見數(shù)字媒體。因此,迫切需要一種技術對動畫進行分類,檢索和過濾。過去幾年中迅速發(fā)展的CBIR技術雖然對靜態(tài)圖像取得了滿意的效果,但是這些技術并不是針對動畫設計的,無法直接用于動畫的分類,檢索和過濾。 鑒于此,本文嘗試在傳統(tǒng)CBIR技術的基礎上,提出一種用于基于內容的動畫短片分類方法。由于現(xiàn)在許多動畫被用作廣告,對用戶來說是一種垃圾信息,因此找到一種檢測和過濾這種信息的方法是很有價值的,在本文中,主要按照這兩類對動畫進行分類�？紤]到動畫與圖片分類的主要不同來自于特征提取,而分類器并不關心其輸入的特征向量是來自于動畫還是圖片,因此本文將重點放在特征的提取和分析上。 本文首先介紹了基于內容的圖像檢索技術的發(fā)展現(xiàn)狀、系統(tǒng)構架以及關鍵技術基礎,詳細闡述了圖像語義特征的提取方法,分析方法以及常用的分類方法,鑒于本文分類目標的特殊性,還介紹了一些其它的特征提取方法,例如圖像中文字區(qū)域的識別等。 在提取特征的基礎上,本文使用互信息量(MI)對不同特征的有效性進行了分析,對提取的不同特征的判別力進行了比較;此外,還分析了將動畫整體考慮與將其看作一系列圖片考慮時的不同,指出后一種做法的效果較差。 最后,本文使用RBF核的支持向量機(SVM)作為分類器,對特征分析的結果進行了驗證,不但比較了單個特征的分類結果,也比較了不同特征的組合的分類結果。最終的分類結果驗證了對特征進行分析時的結論,最后最優(yōu)的特征組合平均錯誤概率達到了8.28%。
[Abstract]:With the rapid development of Internet technology, there are more and more multimedia information on the Internet, especially in the past two years, digital animation, after music and pictures, has become another common digital media used to spread information on the Internet. Therefore, there is an urgent need for a technology to classify, retrieve and filter animation. The rapid development of CBIR technology in the past few years has achieved satisfactory results for still images, but these techniques are not designed for animation and can not be directly used for animation classification, retrieval and filtering. In view of this, based on the traditional CBIR technology, this paper proposes a method for content-based animation short film classification. Now many animations are used as advertisements, which is a kind of junk information for users, so it is very valuable to find a way to detect and filter this information. In this paper, we classify the animation according to these two kinds of animation. Considering that the main difference between animation and image classification comes from feature extraction and the classifier does not care whether the input feature vector is from animation or picture, this paper focuses on feature extraction and analysis. This paper first introduces the development of content-based image retrieval technology, the system framework and the key technical basis, and describes the image semantic feature extraction methods, analysis methods and common classification methods in detail. In view of the particularity of the classification object in this paper, some other feature extraction methods are also introduced, such as the recognition of the text region in the image, and so on. On the basis of feature extraction, the validity of different features is analyzed by using mutual information quantity (MI), and the discriminant power of different features is compared. The difference between considering animation as a whole and considering it as a series of pictures is also analyzed, and it is pointed out that the effect of the latter method is poor. Finally, the support vector machine (SVM) based on RBF kernel is used as the classifier to verify the results of feature analysis, not only comparing the classification results of individual features, but also comparing the classification results of different feature combinations. The final classification results verify the conclusion of the feature analysis, and the average error probability of the optimal feature combination reaches 8.28%.
【學位授予單位】：上海交通大學
【學位級別】：碩士
【學位授予年份】：2007
【分類號】：TP391.4

【引證文獻】

中國碩士學位論文全文數(shù)據(jù)庫 前2條

1 仝琳;論攝影技術在動畫制作中的重要作用[D];山東師范大學;2010年

2 劉永翔;基于支持向量機的瓦斯突出預測研究[D];太原理工大學;2012年

，

本文編號：2291341