利用背景資訊與人臉特徵網路強化側臉人臉屬性偵測

一般使用者所養成的上傳照片習慣，使得網路上的照片(consumer photo)快速增長。如何有效且快速的分析其內容，變成一個很重要的研究方向。現今研究指出。有出現人的照片是最容易讓人印象深刻的照片。為了方便分析這些包含人的照片，各種自動人物屬性偵測的技術如雨後春筍般出現。然而他們幾乎都只針對正臉去做討論，忽略了難度更高且在照片中也大量出現的側臉。此篇論文提出使用順向映射(forward mapping)之方式，來擷取側臉上相對應的區塊。利用此方式，可相當幅度改善基於視覺人臉屬性偵測(visual-based attribute detection)的準確度。除此之外，我們觀察到人臉屬性之間是有關連性存在的(例如老人通常會戴眼鏡)，這代表著可以利用屬性之間的關聯性來幫助我們對於不同人臉屬性間，一起做偵測。同時我們還發現，同一張照片裡不同的人之間的屬性，也存在著關聯性(例如同一張照片裡面的人，大多時候都是同一種族)。因此，我們提出了利用人臉內部和外部關聯性(intra and inter attribute relations)來進一步提升基於視覺上側臉人臉屬性偵測的準確度。實驗結果顯示以上方法確實有效，有助於使側臉屬性偵測這個困難的問題，做得更好。

關鍵字

人臉屬性偵測；側臉；人臉特徵網路

並列摘要

With the rapid growth of consumer photos, efficient and effective image analysis becomes a promising direction for recent researches. Especially, recent researchers found that the most memorable photos usually contain people. Hence, to provide more information for photos with people, automatic attribute detections spring out. Most prior studies focus on detecting attributes for frontal face while it is also essential and more challenging to detect attributes for profile face. We propose an idea of using forward mapping approach to extract corresponding facial parts on profile faces. The proposed method can considerably reduce the error rate of visual-based attribute detection for profile faces. In addition, we observe that facial attributes are highly correlated to each other (e.g., the elder might wear glasses) which implies we can jointly estimate the detection results between different attributes that depicting a person. Moreover, each photo usually contains more than one face. It means that there are some extra contexts information could be utilized since attributes between faces might also be correlated (e.g., family members are usually of the same race). Hence, we propose two context-augmented methods (intra and inter attribute relations) to further improve visual-based profile face attribute detection. Experimental results show that the proposed method can improve the facial attribute detection results indeed for side-viewed faces.