In this paper, we address two problems concerning the contextual information. First, we extend the traditional MRF model into multi-scale structure and provide a frame for dealing the ill-posed problems in multi-scale space. We also apply it to multi-scale image restoration. Second, we provide a new representation and segmentation scheme that is based on the dynamic link architecture neural network by Terman and Wang. We first analysis the traditional MRF model and point out that the limit of the neighbor-hood system caused the defects of the MRF model. It is hard to optimize the objective function and utilize the contextual information with such limited neighbor- hood. We make a new definition of neighbor-hood in scale-space and extend the traditional MRF Model into pyramid structure and hence extend the constrain (such as smooth constrain) into pyramid structure. For a more robust and efficient method, we combine the outlier rejection method in our model. Considers the presence of edge, our method can give a much accurate ourlier detection and hence can provide the edges from being overly smoothed. We also extend the dynamic link architecture neural network to segment images in scale space. In this kind of neural network, the segmentation result is represented by the correlation of the activities of the neurons. The neurons within the same object will tent to synchronize in their oscillation, while the neurons in different object will desynchronize. This new representation scheme resolves the problem of "combination explosion".
修改评论