視覺方向估測技術用於帶景人像構圖計算

近年來，數位相機的普及率越來越高，許多數位相機廠商也開發多種內建功能來協助使用者拍攝照片，但這些功能主要是針對影像的亮度及彩度自動做修正，關於攝影構圖技巧的功能則較少。其實一張好的照片除了亮度、對比與彩度皆要適當調整以外，往往構圖技巧也會影響照片的美感程度。本論文結合人臉偵測與頭部估測技術，發展一套演算法來估測影像中人像所在位置與其視覺方向，以便未來能搭配攝影學中的前置構圖技巧，評估照片是否符合基本構圖法則，或是運用在數位相機上，直接自動偵測並建議最佳取景位置。在本論文中，我們直接將人像的頭部方向當作其視覺方向。首先，使用Viola-Jones 的人臉偵測演算法來偵測人臉區域，接著在偵測到的人臉區域中，分別計算利用眼睛、鼻子與嘴巴間的幾何關係和特徵來偵測其在人臉上的位置，最後利用臉部特徵之間的相對位置來估測頭部方向。在 297 張取自網際網路的測試影像中，人臉偵測的偵測率為 91.9 %；在有偵測到人臉的情況下，頭部估測的正確率則可達到 90.7 %，而每張影像的平均偵測時間約為0.58 秒。

關鍵字

人臉偵測；頭部方向估測；視覺方向估測；帶景人像構圖

並列摘要

In the recent years, digital cameras are becoming more and more popular. Many manufacturers of digital cameras have developed a lot of built-in functions to help users taking a high quality picture. However, those built-in functions mainly perform brightness and chroma automatic adjustment and only few of them are about the photographic composition. In fact, a good picture has not only appropriate brightness, contrast and chroma levels, but also a better composition. In this thesis, we combine techniques of face detection and head pose estimation to develop a method for estimating the position and gaze direction of persons in a picture. The proposed method can be used to verify the foreground space position rule. In addition, it can be implemented as a built-in function of a digital camera to detect and to suggest the best view satisfying popular composition rules. In this thesis, we assume that the head pose direction is approximately aligned with the gaze direction. First, we use Viola-Jones’s face detector to detect the face region in a picture. Then, we use the geometric and texture properties of the eyes, nose and mouth to detect the locations of the facial features. Finally, we use the locations of the facial features to determine the head pose from their relative configuration. In 297 test pictures collected from the Internet, the Viola-Jones method achieved a correct face detection rate of 91.9%. When a face is detected, we used the proposed method to determine the face direction and the accuracy of the head pose estimation is 90.7%. The average computation time is about 0.58 second per image.

並列關鍵字

face detection ； head pose estimation ； gaze direction estimation ； environmental portrait composition

參考文獻

[1] C. T. Shen, J. C. Liu, S. W. Shih, and J. S. Hong, “Towards intelligent photo composition-automatic detection of unintentional dissection lines in environmental portrait photos,” Expert Systems with Applications, vol. 36, no. 5, pp. 9024–9030, 2009.