基于轮廓线段簇的隐式形状模型及其优化方法研究

Shape is an important feature to characterize the class of an object. It is a hot research topic in the domain of object detection to construct an object model by utilizing the relationship among local shape features. To solve the problem of mis-match between shape features of complex images, by using the basic idea of Hough transformation, this project proposes a method of compact feature representation and combination based on contour-line-segment-cluster, giving its similarity measurement and learning approach. To reduce the impact of feature distribution and amount on the feature combination, we propose a method of building an implicit shape model based on two-level multi-objective optimization. To improve the generalization capacity of such implicit shape model, during the shape matching in a sliding window, we introduce a paired matrix that supports partial features matching, and give a method of shape matching based on a graduated assignment algorithm. Then, we obtain a global model transformation and a matching relationship between local features. To improve efficiency and effectiveness, the parameters of the sliding window are determined by a multi-objective optimization method based on bipolar-preference. We would like to perform detection experiments on complicated datasets such as TU Darmstadt to confirm the effectiveness of the proposed models. The project provides a completely new thread for the line-segment representation of shape features and the construction of implicit shape model, which enlarges the applications of contour-line-segment features to complex images. The research has extensive application prospects in national defense, security check, and so on.

形状是刻画物体类别的重要特征，利用局部形状特征之间的关系来构建物体模型的方法已成为物体检测领域中的研究热点。针对复杂图像上形状特征误匹配问题，本项目利用霍夫变换思想，提出一种紧凑的基于轮廓线段簇的特征表示及组合方法，给出该方法的相似度度量及其学习方法；为了减弱线段分布和数量对线段特征组合的影响，提出一种基于双层多目标优化的隐式形状模型构建方法；为提高该隐式形状模型的泛化能力，在滑动窗口形状匹配中引入支持部分特征匹配的配对矩阵，给出基于渐进分配的形状匹配方法，求得全局模型变换和局部特征配对关系，提出融合双极偏好的多目标优化方法进行滑动窗口参数的选取，提高匹配效率和效果。通过TU Darmstadt等复杂数据集上的物体检测实验，验证该模型的有效性。本项目为形状特征的线段表示和隐式形状模型的构建提供了一种全新思路，扩大了轮廓线段特征在复杂图像上的应用，该方法在国防、安检等领域具有广泛应用前景。

本项目对基于轮廓线段簇的隐式形状模型进行了系统而又深入地研究，提出了基于霍夫变换思想的轮廓线段特征的优化方法，解决了若干实际应用问题。遵循项目申请的研究路线，项目组提出基于轮廓形状信息的局部线段特征表示，设计相应的相似度测量函数及度量学习方法。通过以检测率和误检率为目标，提出一种基于双极偏好的滑动窗口参数优化方法，并通过引入决策者对两个指标的正负偏好，使置信度阈值的选择更加合理，从而提高隐式形状模型的物体检测效果。引入多目标进化算法，提出一种基于双层多目标优化的隐式形状模型构建方法，利用多目标进化算法随机组合策略和全局搜索能力在减少特征数量的同时获取较好泛化能力的物体隐式形状模型。对不确定偏好下种群的进化引导机制及大规模变量下多目标优化问题的求解进行了深入研究，设计出偏好向量生成策略，消除算法收敛性对决策者偏好的敏感性，提出APD精英选择策略，进一步提高多目标进化算法对隐式形状模型的物体检测精度；提出了大规模变量分解的多目标粒子群优化算法，通过变量随机分解策略来增加关联变量分配到同组的概率，有效地提高了算法的运行效率，进而在多目标优化的基础上解决隐式形状模型的优化问题。本项目还研究解决了高维目标优化及大规模变量优化中的若干关键问题，所提方法已被用于解决供应链管理、工程设计、系统建模和图像处理等领域中的优化问题，具有较强的科学意义和广泛的应用价值。