Research On Holon Representation Of Large-Scale Image Semantic Segmentation

In this thesis,research on holon representation of large-scale image semantic segmentation based on latent semantic inference approach and multi-scale deconvolution mechanism is presentated.The research objective is to develop segmentation models and corresponding algorithms to capture the vital information of natural images,and finally,to achieve pixel prediction and semantic segmentation within a structural architecture.The proposed architecture consists of three main aspects: latent semantic inference for natural image segmentation,multi-scale deconvolution network for large-scale image semantic segmentation,and further interactive and unsupervised object segmentation from public resources,i.e.,Youku and Youtube.Latent semantic inference apts to exploit the characteristics of local pixels and global regions in a given natural image for enhancing image segmentation in a disciplined manner,which overcomes the limitation of homogeneous superpixel-based treatment.The proposed approach foucus on the crucial point of how to integrate the multiple information into the segmentation process by coupling potential functions defined on labels with the designed Cross-Region and Cross-Scale potentials.Multi-scale deconvolution network makes it possible to train a single net for joint detection and semantic segmentation task with a decent speed.The proposed network introduce multi-scale deconvolution mechanism,which is composed of down-scale and up-scale stream,to combine the multi-scale features instead of the using the multi-scale inputs which has been demonstrated that outperforms average-pooling and max-pooling in a network.Further video object sementation is achieved in an interactive and unsupervised manner.By drawing simple lines to prompt the information of the interest object in the first frame,the proposed method can automatically perform the Gaussian Mixture Model(GMM)labeling which can get the interest object class,then parameters updating,and Graph Cut performing,to segment the interest obect in the video.