歐陽明Ouhyoung, Ming臺灣大學:資訊工程學研究所吳偉成Ng, Wai-SengWai-SengNg2010-06-092018-07-052010-06-092018-07-052009U0001-2507200915583500http://ntur.lib.ntu.edu.tw//handle/246246/185427由於數位攝影的普及，現在拍照和保存照片可以成為一件容易的事。然而，並非很多人都熟悉照片的美學規則，如構圖和顏色分配。為一個工具，我們會介紹在傳統攝影學上，如何對構圖做一些量化析，一些情況例如水平線，照片平衡，主體位置，線條及形狀和人與背景線穿透造成融合。每個規則的加權指數是實驗後才確定的，含來自Flickr網站500張照片和幾十個主題。支援向量迴歸演算法被首次使用來量化和“預測”人的喜好程度，然後，1萬多張來自Flickr的照片可再用於最後使用者分析。使用者分析與實驗證實了我們最初的假設，即自動照片排名是有效的。此外，我們總共用了2000張照片(dpchallenge網站)分別用於學習（ 1000 ）和測試（ 1000 ），並將這兩個類別分別出好的或不好的照片。可在上述實驗能夠獲得80.9%的準確度，比之前別人的結果好(3000張/72%, 2400張/74%, 1800張/76%)。Due to the popularity of digital photography, taking, viewing, and preserving photos are much easier. Nevertheless, not many people are familiar with photo esthetics rules, such as composition and color distribution. As an tool, we introduce a quantitative analysis method of photo composition based on well-known photography rules, such as horizon balance, intensity balance, locations of region-of-interests (ROIs), line patterns and merger avoidance. The weighting factors for each of the rules are determined by an experiment involving 500 photos from Flickr sites and dozens of subjects. Support Vector Regression techniques are first used to quantify and ”predict” human evaluation, then results from more than 10,000 photos from Flickr are used for theinal user study. The user study experiment corroborates with our initial hypothesis that automatic photo ranking is effective. Furthermore, 2000 photos from dpchallenge website are used for training (1000) and test (1000), withwo class classification to find the better and worse ones, and we are able to get 80.9% accuracy as compared to human rating.致謝i要iiibstract iv Introduction 1 Related Work 5.1 Automatic image cropping . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Two-class classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Feature detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Learning of dataset approach . . . . . . . . . . . . . . . . . . . . . . . . 7.5 Restrictions of evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6 Information theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 System Overview 11.1 Canny Edge Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Face Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3 ROI Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4 Feature Descritpor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.5 Training and Testing Phase . . . . . . . . . . . . . . . . . . . . . . . . . 15 Rules of Esthetics in Photo Composition 16.1 Photo Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16.2 Color and Intensity Distribution . . . . . . . . . . . . . . . . . . . . . . 24.3 Accuracy of Individual Features . . . . . . . . . . . . . . . . . . . . . . 29 Automatic Ranking of Photos 30.1 Precision and Recall . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 User Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33.3 Speed Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Conclusions and Future work 41ibliography 4329893976 bytesapplication/pdfen-US照片排序照片構圖美學規則顏色分佈Image rankingphoto compositionphotography estheticsthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/185427/1/ntu-98-R96922038-1.pdf