面向复杂场景图像分割的多模态驱动型水平集方法研究

Due to the influences of different factors, it is relatively difficult to segment the target of interest in complex scene. For the scenes where there are some interferential objects around the target, the scene where the serous intensity inhomogeneity phenomenon appears, and the scene where the target presents inhomogeneious texture structure, the incorporation of multi-modal informaton and construction of energy functional in level set method shall be studied in this project. Firstly, the image saliency is extracted as the constraint information. The edge-based level set method shall be designed to incorporate the saliency information, thus constructing the saliency-based energy term. Secondly, the multi-scale inverse bias field function on neighborhoods of each pixel shall be designed with the competitive weighting factor being defined. Then, the multi-scale intensity-based energy term is constructed by adopting the region-based level set framework. Thirdly, the semi-local operator is constructed so that the metric tensor on texture manifold can be defined. Then, the energy term can be constructed by incorporating the semi-local texture feature based on certain similarity measure of texture probability distribution. Finally, the multi-modal property of complex scene shall be analyzed. Correspondingly, the multi-modal driven energy functional can be constructed by combine the energy terms in a weighting way. Then, the level set function shall be represented in a continuous way, thus producing a complete level set segmentation framework. It can make the evolving curve approaches the target of complex scenes under the combinational force of saliency, multi-scale intensity and semi-local texture information, etc. The development of this project can promote the theoretical researches and practical applications of level set segmentation and image processing.

受各种因素影响,在复杂场景中分割出感兴趣目标一般较为困难。针对目标周围存在着干扰物、图像中出现严重灰度不均匀现象、目标呈现出不均匀纹理等场景，本项目将研究水平集方法中多模态信息的融入和能量泛函的构建。首先，提取图像显著性作为约束信息，设计边缘型水平集以融入显著性信息，从而构建显著性能量项；然后，为像素点设计其邻域上的多尺度逆偏置场函数，定义竞争性权重因子，基于区域型水平集构建多尺度灰度能量项；接下来，构造半局部算子并定义纹理流形的度量张量，设计纹理概率分布相似性度量，融入半局部纹理特征以构建能量项；最后，分析复杂场景的多模态特性，对能量项进行加权组合以构建多模态驱动型能量泛函，将水平集函数进行连续性表达，从而形成一个完整的水平集分割框架，使得演化曲线在显著性、多尺度灰度和半局部纹理等信息的联合作用下逼近复杂场景中的目标。本项目的开展将对水平集分割和图像处理的理论研究和实际应用起到推动作用。

水平集方法能处理复杂拓扑结构变化且具有较高的分割精度，可以满足复杂场景图像分割的要求。针对目标周围存在干扰物、图像中出现严重灰度不均匀现象、目标呈现出不均匀纹理等场景，本项目研究水平集方法中多模态特征的融入和能量泛函的构建。. 设计优化种子法以提取图像高阶统计显著性特征；在图像上执行非线性扩散并开展局部分段光滑搜索，从而建立非线性扩散能量项；构造分数阶差分掩码，形成分数阶差分能量项。在显著性能量项的驱动下，水平集演化得以在显著性特征空间上进行，实现了对具有强噪声、干扰物的复杂场景图像的有效分割。. 构建局部最大描述差异特征，设计竞争性权重因子，形成多尺度能量项；抽取局部灰度变异信息，构造基于局部增强区域的多尺度拟合项；基于多尺度逆偏置场函数，提出水平集扫描优化模型。所构建的多尺度能量项可以有效规避局部邻域尺寸的影响，通过在多尺度特征空间上执行水平集演化，可以分割严重灰度不均匀图像。. 在图像块的基础上定义纹理流形的度量张量，得到半局部纹理特征；计算曲线内外半局部纹理特征的概率分布，基于相似性度量构建半局部纹理能量项。半局部纹理特征的融入可以避免Gabor特征空间的高维度计算，且同时包括纹理特征和其空间关系，所形成的能量项对于纹理不均匀图像分割具有较好的鲁棒性。. 构建分块均值-Gabor、局部时空兴趣点-全局傅里叶描述子、HOG-时空兴趣点-骨架结构、熵Canny算子-局部和全局光流等多模态特征；设计基于双重约束松弛、余弦拟合能量、交叉熵等多种水平集模型；进一步从能量泛函相似性度量的角度，对区域型水平集方法进行了系统性研究，归纳为六种代表类型并给出了对应的能量表达结构。. 提出两类基于连续非分段函数的新型距离正则项，设计基于参数Newton型和Thiele型两种连续分数有理插值方法，实现了水平集函数的连续性表达。引入多模态特征权重因子，构建多模态驱动型水平集分割框架，在显著性、多尺度灰度和半局部纹理等特征的联合驱动下分割复杂场景中的目标。. 本项目已实现全部计划目标，相关研究成果为水平集图像分割的理论研究提供了若干有价值的工具，并可被应用于其它图像处理领域，所提出的面向复杂场景图像的水平集分割技术可广泛应用于工农业生产自动化、生物特征识别、医学图像分析、智能安全监控等领域。