Spectral segmentation based on the weighted histogram

摘要

Spectral segmentation algorithms can extract the global impression of an image and be widely used in many areas and applications related to image segmentation. Traditional spectral algorithms need to construct an affinity matrix based on all the pixels in the image and compute the eigenvectors of the matrix to find global optimum segmentation. With the growing of image size, the spatial and computational complexity increases egregiously. We propose a novel approach for solving the problem. Rather than focusing on pixels in the image, our approach aims at the gray histogram extracting the global impression of an image. We construct an affinity matrix directly based on the gray histogram and weight matrix with a histogram coefficient of corresponding gray level. Our method decreases the affinity matrix dimension to 256 × 256 at most for the gray image, and thus decreases the complexity sharply. The spatial and computational complexity is insensitive to the image size. We have used the proposed method to segment text and PCB images, and found the segmentation results to be very encouraging.