透過您的圖書館登入
IP:3.145.156.250
  • 學位論文

利用各種共現特徵來增進人臉辨識效能

The Use of Various Statistical Co-occurrence Patterns for Boosting Face Tagging

指導教授 : 鄭卜壬

摘要


隨著行動裝置的發達,照片的數量也隨之急速提升,如何自動化人 臉標籤也成為一個重要的議題,最常見的做法微使用人臉辨識系統, 但現今人臉辨識技術有其侷限,在各種不利因素下無法正確辨認出使 用者,造成只有部分標籤的情況。在只有部分標記的情況下,本研究 提出了利用人與人間的共同出現的行為模式來增進人臉標記。首先我 們對資料做了詳細觀察,並提取出使用者之間共同出現的特徵,利用 這些特徵建構出半監督式的機率模型,此機率模型能夠根據已知的部 分標籤去推薦剩餘的未知標籤之候選人。我們所提出的模行能夠建置 在人臉辨識系統所產出的部分標籤的結果,且沒有使用任何影像辨識 技術。最後,無論在資料有受到控制(Facebook) 與沒有控制(Picasa) 的 環境下進行實驗,結果證明在兩種情況下,此方法的效能都高於常見 的現行方法。

並列摘要


Face recognition benefits many applications such as photo searching and annotation. However, such technique is not robust enough to identify people under widely varying photo conditions. In this paper, we propose a semisupervised probabilistic graphical model for boosting face recognition by predicting the names of the unrecognized faces in a statistical manner. Different from pairwise co-occurrence context adopted by conventional approaches, we explore more useful context such as the co-occurrence between subgroups, the different frequency of occurrence for people, and the different numbers of people appearing in photos, and integrate them into our model. Experiments on Facebook and Picasa demonstrate that the proposed model can effectively improve the recognition performance, compared to both of unsupervised and supervised learning methods.

參考文獻


recognition in personal photo albums. In Computer Vision and Pattern Recognition,
Symposium on Electronic Imaging Science and Technology. Press, 2006.
Cao. Link prediction and recommendation across heterogeneous social networks. In
[4] A.C. Gallagher and Tsuhan Chen. Using group prior to identify people in consumer
[5] A.C. Gallagher and Tsuhan Chen. Clothing cosegmentation for recognizing people.

延伸閱讀