Image Segmentation Using Multi Dimensional Co-occurrence Matrix (3rd Report)
Tomotaka Sugishita, Terumoto Komori, Yoshihiko Nomura, Norihiko Kato · The proceedings of the JSME annual meeting · 2002
The image segmentation method based on the multi dimensional co-occurrence matrix belongs to the class of region-based techniques. It uses a feature vector composed of multiple features for multiple windows neighboring a pixel. And a mixture Gaussian distribution is fitted to the feature vector set for pixels. There are two problems on dealing with boundary pixels. As for the first problem, boundary pixels cause an ill effect in the estimation of distribution model. It was solved by the exclusion of boundary pixels. The other problem is in a difficulty when assigning the boundary pixels to the estimated mixture Gaussian distribution model. The later problem was examined in this paper. Each pixel is assigned according to the posteriori occurrence probabilities calculated from the features of each pixel.