This thesis addresses the problem of autonomous object segmentation. To do so the proposed segementation method uses some prior information, namely that the image to be segmented will have a low depth of field and that the object of interest will be more in focus than the background. To differentiate the object from the background scene, a multiscale wavelet based assessment is proposed. The focus assessment is used to generate a focus intensity map, and a sparse fields level set implementation of active contours is used to segment the object of interest. The initial contour is generated using a grid based technique. The method is extended to segment low depth of field video sequences with each successive initialisation for the active contours generated from the binary dilation of the previous frame's segmentation. Experimental results show good segmentations can be achieved with a variety of different images, video sequences, and objects, with no user interaction or input. The method is applied to two different areas. In the first the segmentations are used to automatically generate trimaps for use with matting algorithms. In the second, the method is used as part of a shape from silhouettes 3D object reconstruction system, replacing the need for a constrained background when generating silhouettes. In addition, not using a thresholding to perform the silhouette segmentation allows for objects with dark components or areas to be segmented accurately. Some examples of 3D models generated using silhouettes are shown.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:589883 |
| Date | January 2013 |
| Creators | McDonnell, Ian |
| Publisher | University of Warwick |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://wrap.warwick.ac.uk/58630/ |
Page generated in 0.0015 seconds