A Research On Image Segmentation And Dimensionality Reduction

In the first part of this paper, we present a new Markov Random Field based FCM image segmentation algorithm. A new energy function is proposed to utilize the spatial and contextual information simultaneously. In the proposed energy function, we use a weighted distance to reflect the different effects of neighborhood pixels. By using the new energy function, the new algorithm has a better performance in noise-corrupted images. Experimental results on real and synthetic images show our method is effective.In the second part of this paper, we analyze the relationship between selecting bands from a hyperspectral image and choosing meaningful columns from an arbitrary matrix. Inspired by the analogy, this letter proposes a new band selection algorithm based on column subset selection. Considering the high dimensionality of each band, we improve the algorithm to obtain a higher selection quality. The quality is assessed by applying supervised classification on the selected bands. Compared with three other unsupervised methods, experimental results show our approach, named Band Column Selection(BCS), has a competitive performance. We demonstrate that BCS is robust to noisy bands. Good classification accuracy on original and refined data(noisy bands are removed) are both obtained.In the third part of this paper, we deeply investigate the choice of distance measure in the proposed BCS method. First of all, we introduce the CUR decomposition to evaluate the effectiveness of band selection. Then, we apply our BCS on a decomposition-rebuild benchmark to see which distance measure is better. Experimental results show that when the selected column is short or the dimensionality is low, BCS with Manhattan distance measure preforms better. However, when the dimensionality is high, distance measure should pick a higher norm to maintain the effectiveness of BCS method.