Domain conversion with local posteriors for image segmentation
EunSang Bak, Kayvan Najarian · 2004
The estimates of the posterior probabilities of the attributes in the image are widely used as criteria for image segmentation. The methods using this measure, however, suffer from intrinsic errors that occur around the boundary between regions. The errors are caused by estimating the posterior probabilities over the entire image. To resolve this problem, we define novel local posterior probabilities to better capture the local characteristics and then use them in an iterative segmentation process. Furthermore, the image itself is converted to another image in a new domain by a domain conversion method. It is shown that the converted image in the new domain is less susceptible to intrinsic errors.