Posture Guided Segmentation For Human Image

As the number of picture in the Internet strictly increasing and the easy access to image capturing device, more and more situations need to handle kinds of problem of image processing. Image segmentation is an important filed of computer vision, and the research filed to segment fixed class of objects in image gets more and more attention. Semantic segmentation which combined image segmentation and object recognition has lots of applications.Semantic segmentation segments some fixed objects from the background of image by taking advantages of global semantic features in image. Different from traditional image segmentation, semantic segmentation stress the affection of semantic features in conducting the whole segmenting process. This paper provides a new method to segment persons from background by combining posture of human detected from a special detector with some local features. To preserve contours of person, we over segments the image. A graph of all nodes represented by segmented superpixels is constructed. The weights of all edges are computed as the ratio of intersection to union between superpixels and pose mask. At the same time, another graph is constructed based on the image local features. Finally the two graphs are used as inputs of a multi-view spectral cluster model, and then all pixels are divided into person and background.