Unified Context Feature Extraction For Image Segmentation

Image segmentation,including motion segmentation,general semantic segmentation,and human parsing(human body part segmentation),is a key technology in the field of computer vision,and has important application value in various fields such as video surveillance,intelligent driving,and human-computer interaction.In pixel classification-based image segmentation tasks(also known as semantic segmentation),the contextual information of the pixels plays a decisive role in pixel-wise classification.Existing mainstream methods separate context feature extraction from tasks such as classification and resolution recovering.On the one hand,this separation strategy is difficult to achieve the optimal overall segmentation effect.On the other hand,it also seriously reduces the efficiency of segmentation.In response to the above problems,this paper proposes a set of unified context feature extraction methods,which simultaneously improves the effect and efficiency of segmentation.According to the semantic order of the segmentation results from low to high and from coarse to fine,the work and main innovations of this thesis are as follows:(1)For the single image motion region segmentation task,a set of unified methods for blur context feature extraction and classification are proposed.The separability of the novel motion blur indicator function determined by the local gradient context feature is proved in detail and an intuitive geometric explanation is given.According to the geometric explanation,a single-class classifier for motion blur kernel classification determined by the angle threshold is designed,and achieves efficient and high-precision motion blur region segmentation.(2)For the general image semantic segmentation task,an image semantic segmentation method combining unified context feature extraction and resolution recovering with feature-guided filtering is proposed.In order to further improve the integration degree and segmentation efficiency,an integrated context feature extraction and resolution recovering method based on content adaptive fusion is proposed.The proposed methods overcome the shortcomings of low efficiency and poor accuracy of the methods adopting separation strategies for context feature extraction and resolution recovering,and at the same time significantly improves the efficiency and accuracy of image semantic segmentation.(3)For the fine-grained semantic human parsing task,a hierarchical human parsing method based on typed part-relation reasoning is proposed: for part decomposition relation,composition relation at different levels and part dependency relation at the same level,three different and type-adaptive relation network modules are developed respectively,then an iterative message-passing network based on spatial convolution and typed edges is constructed to take full advantage of the loopy structure in the human part topology.It solves the problems of incompletion and weak adaptability in relation modeling of traditional methods,and significantly improves the accuracy of human parsing.The above-mentioned unified context feature extraction method realizes highefficiency and high-precision motion segmentation,general semantic segmentation and human parsing,and has broad application prospects in the fields of intelligent surveillance,autonomous driving,and human-computer interaction.