Robust Eye Biometry and Genetics | Stine Harder, Susanne Rytter Christoffersen
| | Abstract | The goal of this thesis is to make an objective evaluation of the human iris colour and structure, in order to extract quantitative measures for DNA correlation. In order to ful lfl this goal the iris region was extracted from high resolution eye images provided by the Department of Forensic Medicine, Section of Forensic Genetics. Inspired by a method proposed by John G. Daugman we have succeeded in extracting the iris region. Furthermore the eyelid boundaries were located using a pixel classi fication with a Markov Random Field approach. We were able to locate the pupil boundary and iris boundary correctly in 75:0% and 96:4% of the cases and the upper and lower eyelid boundaries in 76:6% and 73:0% of the cases, respectively. The total percentage of correctly extracted irises is 64:3% using a Dice Coe fficient of 0.92 for comparison with human annotations. The evaluation of human eye colour was performed by generating a colour classifi er, an explanatory ratio, and by performing a colour based image clustering. The colour classi er was trained on subjective evaluations of human eye colour, and the percentage of correctly classi fied images was 65:4%. We have calculated a ratio explaining the amount of blue and brown coloured regions in the iris, which has shown to be a promising continuous measure for iris colour since it clearly distinguishes between blue and brown eye colour. The colour based image clustering is a data driven approach, where the division into colour groups are very robust compared to a subjective evaluation. The image groups generated follow the general perception of eye colour being divided into groups ranging from blue, through intermediate, to brown. This is in agreement with the knowledge about DNA. A minor study of human perception of eye structures was conducted, resulting in Fleiss' Kappa values between 0.23 to 0.56 for different structures. A structure based image clustering was performed in order to generate a data driven evaluation of iris structures. The result is a division into three groups, one consisting of very smooth eye images, one consisting of eye images with a high amount of structure and one consisting of images with a large re flection area. This division is very promising regarding a DNA correlation. However some images were incorrectly clustered, which makes us question the result. | | Type | Master's thesis [Academic thesis] | | Year | 2012 | | Publisher | Technical University of Denmark, DTU Informatics, E-mail: reception@imm.dtu.dk | | Address | Asmussens Alle, Building 305, DK-2800 Kgs. Lyngby, Denmark | | Series | IMM-M.Sc.-2012-08 | | Note | Supervised by Associate Professor Rasmus Reinhold Paulsen, rrp@imm.dtu.dk, DTU Informatics | | Publication link | http://www.imm.dtu.dk/documents/ftp/ep2012/ep12_08-net.pdf | | BibTeX data | [bibtex] | | IMM Group(s) | Image Analysis & Computer Graphics |
|