以美學為基礎的照片構圖量化分析

Abstract

數位相機的普及率越來越高，因此越來越多人拍照，使得數位照片成 指數性成長。面對龐大數量的照片，照片的整理與挑選成了一個很大 的問題。 照片構圖的意思是相機位置的擺設以及視野的選擇，決定什麼物體 該放進來或該排除在外，一般來說，照片構圖幾乎決定了一張照片的 好壞，所以我們針對照片構圖的部分做研究。照片構圖這個問題往往 是主觀的，也跟人類的視覺感受有關，雖然是主觀的，但幸運的是， 根據攝影學家多年來的經驗，可以萃取出一些比較通用性的原則，這 些規則在各種談攝影的書中被談到。我們從書中挑選出幾個定義較明 確、且較適用於電腦量化分析的規則，將他們實作成可以自動化執行 的程序，用這些規則挑選出比較不好的照片，以及對照片做一些建議 與評分。 一張不是很好的照片如果經過很好的裁切，可以讓一張照片起死 回生，現在數位相機的能力也越來越強，照片的解析度越來越高，在 沒有輸出太大尺寸照片需求的原則下，很有本錢可以對照片做一些裁 切，我們以構圖分析的結果來做裁切的原則，有機會能讓照片變的更 好。 我們初步的結果是基於132 張照片，這些照片是針對每個規則（水 平線、照片平衡、主體位置、線條及形狀、避免融合）挑選出比較 簡化的情況，每個規則的準確率分別在71%、96.8%、73.1%、78.4%以 及71.4%。

Digital camera is very popular and the number of digital photos grows exponentially. Faced to huge number of photos, the collection and selection of photos becomes a big problem. Photo composition means the placement of camera and the selection of the field of view. It determines whether objects should be placed inside the photo or be excluded outside. Most people agree that photo composition almost determines whether a photo is good or not. Therefore, our research is focus on photo composition. This problem is subjective and relates to human visual perception. Although this problem is subjective, it is fortunate that from the experience accumulated by photographers in recent years, certain common rules were extracted. We select the rules that have clear definition and suitable for automatic quantitative analysis to pick out bad photos or to make recommendations and scores. It is said that a not very good photo after doing a good cropping will let a photo turn from death into life. The performance of digital camera nowadays is more and more powerful and image resolution is higher and higher, and so images can be cropped without serious loss of resolution. Our results in automatic cropping support this observation. Our preliminary results for 132 photos are based on five rules (the horizon, photo balance, location of main object, revealing line patterns and shapes, and avoiding mergers), and the precision of each rule is 71%, 96.8%, 73.1%, 78.4%, and 71.4%, respectively.