請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74327
標題: | 針對在多樣光照下進行車內人臉辨識設計基於三元組概念之訓練架構 A Triplet-based Training Architecture for In-vehicle Face Detection under Various Lighting Situations |
作者: | Yi-Chen Chuang 莊宜蓁 |
指導教授: | 莊永裕(Yung-Yu Chuang) |
關鍵字: | 人臉偵測,車內人臉偵測,影像增強,正規化, Face Detection,In-vehicle Face Detection,Image Enhancement,Normalization, |
出版年 : | 2021 |
學位: | 碩士 |
摘要: | 近年來隨著智慧型汽車相關技術的成熟,車內人臉辨識逐漸受到重視。然而一般的人臉偵測模型尚無法在車內人臉辨識上獲得較好的結果。其原因我們推測為車內人臉辨識會碰到較多在不同光照下的例子,而一般的人臉偵測模型著重於大多數情境下的表現結果,經常忽略在較極端光照下的例子所致。在本論文中我們試著消除不同光照對圖片造成的影響。我們提出一個訓練架構來訓練出能夠對輸入資料在偵測前進行正規化處理的正規器,並將此正規器和人臉偵測器接在一起進行端對端訓練優化。我們的方法在基線做得最差的測試情境中相比基線結果進步了 47.27%。 As intelligent vehicle technologies become mature, in-car face detection gradually draws everyone's attention. However, general face detection models have yet to perform good when it comes to in-car face detection. We guess the reason is that while in-car face detection has to deal with more cases from various lighting situations, general face detection tends to focus on performing good on major cases and often ignores cases under extreme lighting situations. In this thesis, we tried to remove the effect lighting had on images. We proposed a training architecture to train a normalizer to normalize input images before getting detected, jointed it with a face detector and did end-to-end training for optimization. Our method outperformed baseline by 47.27% in the test scenario that baseline performed worst on. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74327 |
DOI: | 10.6342/NTU202100125 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2201202113553400.pdf 目前未授權公開取用 | 22.5 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。