Multisample rank-based image segmentation
Roman V. Podrezov · Proceedings of the Russian higher school Academy of sciences · 2015
This paper considers the image segmentation task in a prior uncertainty conditions.Image observations are assumed to be independent.The rank-based segmentation algorithms are appropriate under these conditions, but a drawback of estimating the threshold based on a single sample with a fixed position or on adjusting the sample position is high sensitivity to the position of the class area and the sample.In contrast to the existing rank methods the proposed method uses a maximum likelihood criterion which is calculated by multiple samples.An effective algorithm of calculating the decision statistics is proposed.The algorithm is to match each rank to the sample number and to use the sample number for evaluating the log-likelihood function increment.The algorithm was tested on image models with normally distributed classes.Performance characteristics are provided confirming a possibility to solve gray-level image segmentation tasks.