bachelor thesis

SignatureBA 018
AuthorSophie Paul
TitleImplementation and Examination of an Alpha Matting
TutorDr.-Ing. Matthias Kunter, Dr.-Ing. Sebastian Knorr
ProfessorDr.-Ing. Thomas Sikora
AbstractThe subject of this thesis is the implementation and examination of a chosen alpha matting algorithm. Alpha matting in general is a technique from the field of computer vision and computer graphics and deals with the extraction of a foreground object from its background. At this it is similar to the more popular image segmentation which deals with the separation of different parts in an image. Image segmentation creates segments with hard borders. In contrast the here addressed alpha matting tries to define the boundaries of foreground objects with fuzzy outline. Therefore it computes the opacity of relevant pixels.
After introducing the topic alpha matting, a description of exemplarily selected alpha matting approaches is given. Afterwards the chosen Quick Matting Algorithm as well as an extended version of this approach is specified. The implementation is realized in an existing C++ framework.
Afterwards the matting results are evaluated using different test images with corresponding ground-truth mattes as well as an image sequence. A quantification of the chosen algorithm is done based on error measurements and visual comparison. At this the chosen algorithm shows good performance for images with smooth and compact foreground regions. It fails if the images contain thin and highly transparent foreground regions as well as for videos. The extended version of the algorithm shows improved performance. Finally an outlook on possible other improvements is given.
Key wordsalpha matting algorithm, computer vision, computer graphics, Quick Matting Algorithm, C++ framework, implementation