Locating Landmarks Using Templates
Robustní metody a regrese
rigorózní práce (UZNÁNO)
Zobrazit/ otevřít
Trvalý odkaz
http://hdl.handle.net/20.500.11956/14314Identifikátory
SIS: 57186
Kolekce
- Kvalifikační práce [10690]
Autor
Fakulta / součást
Matematicko-fyzikální fakulta
Obor
Pravděpodobnost, matematická statistika a ekonometrie
Katedra / ústav / klinika
Katedra pravděpodobnosti a matematické statistiky
Datum obhajoby
25. 2. 2008
Nakladatel
Univerzita Karlova, Matematicko-fyzikální fakultaJazyk
Angličtina
Známka
Uznáno
This thesis is devoted to automatic location of landmarks (month and eyes) in images of faces using templates. There is an unsatisfactory experience with existing software because of its high sensitivity to small rotations of the face. The weighted correlation coefficient as a similarity measure between the template and the image turns out toout- perform the classical correlation. It is presented how to choose the weights to increase the discrimination of the parts of the face which correspond to the template from those which do not. Optimization without constraints tends to degenerate and to obtain a ro- bust version we bound the influence of single pixels. In a similar way the template can be optimized to improve the discrimination further. The results arc? compared for different initial choices of weights and their robustness to different size or rotation of the face is examined. The method docs not use any special properties of faces and can be classified as a robust nonparametric disrimination technique. Abstrakt Pra.ce HC zabyva, automatickym liledam'm objcktn (list a oci) v obrazech oblicejri za po- rnoci sabloii. Dostupny software je velmi citlivy k malym otocenim oblieje a zkusenost s mm je neuspokojiva. Vaznny korelarm koefieieiit je vhodnejsi mi'ron podobnosti mezi sablonou a obrazem nez...
This thesis is devoted to automatic location of landmarks (month and eyes) in images of faces using templates. There is an unsatisfactory experience with existing software because of its high sensitivity to small rotations of the face. The weighted correlation coefficient as a similarity measure between the template and the image turns out toout- perform the classical correlation. It is presented how to choose the weights to increase the discrimination of the parts of the face which correspond to the template from those which do not. Optimization without constraints tends to degenerate and to obtain a ro- bust version we bound the influence of single pixels. In a similar way the template can be optimized to improve the discrimination further. The results arc? compared for different initial choices of weights and their robustness to different size or rotation of the face is examined. The method docs not use any special properties of faces and can be classified as a robust nonparametric disrimination technique. Abstrakt Pra.ce HC zabyva, automatickym liledam'm objcktn (list a oci) v obrazech oblicejri za po- rnoci sabloii. Dostupny software je velmi citlivy k malym otocenim oblieje a zkusenost s mm je neuspokojiva. Vaznny korelarm koefieieiit je vhodnejsi mi'ron podobnosti mezi sablonou a obrazem nez...