图像语义目标分割旨在建立图像像素与物理目标之间的关联，从而消除人对图像的高级语义理解与计算机的低级视觉表达之间的差异，是实现图像智能分析与理解的核心关键技术。本项目拟以目标的形状语义及形状上下文为突破口，辅以图像纹理等底层特征，基于变分法、图模型、最优化等研究手段，在水平集框架下通过自下而上和自上而下两种模式实现语义目标分割与识别。本项目的研究内容有：1）通过流形、李群及非负矩阵分解等方法，研究目标形状的对齐以及形状语义的提取与建模方法；2）基于图模型，研究目标形状上下文语义的表示及推理学习方法；3）提升水平集曲线演化框架，通过统一的能量函数整合低层视觉表达与上下文语义信息，实现具有形状上下文约束的多语义目标分割与识别。以上研究有助于完善和发展基于水平集的图像分割理论与方法，为图像分析与理解等研究领域提供新思路与新方法。

Aiming at mapping image pixels to the corresponding objects, semantic object segmentation of images bridges the gap between the high-level semantic understandings of human and the low-level visual express of computers, and it is the key technique for the implementation of intelligent image analysis and understanding. Based on the methodologies, i.e., calculus of variations, graphical models and optimization techniques, our application attempts to integrate shape semantics, shape context as well as texture characters to realize semantic object segmentation in the two manners of bottom-up and top-down. The main contents of our application include: 1) researching the methods of shape alignment and extraction of shape semantics by manifold, Lee group and non-negative matrix factorization; 2) based on graphical models, researching the techniques of inference and learning of shape context models in scenes; 3) improving the level set framework for contours evolution to integrate low-level visual clues and high-level shape context into a unified energy functional, and finally to realize the segmentation and recognition of multiple semantic objects constrained by multiple semantic contexts. Our application would greatly contribute to consummating and developing the theorems and methodologies of level set based segmentation, and provides image analysis and understanding with new thoughts and techniques.