倾斜影像高效匹配的近邻信息监督度量学习方法

Recently aerial oblique images have re-attracted the public’s attention due to its abilities to view all the facades of the buildings and to be easily interpreted. However, the penta-view camera systems have leaded to tremendous pair-wise stereo matches and the inefficiency of the high dimensional SIFT-like descriptor has impeded applications using the matching paradigm with SIFT-like feature descriptors. Dimension reduction is intuitively a possible solution, but traditional techniques for this matter are generally relied on the Gaussian assumption of the data. This proposal formulate the dimension reduction of feature descriptor as a large scale supervised metric learning problem. With the labeled information of training samples, the proposed method does not require on any distribution assumptions. Dimension reduction is achieved by low-rank matrix decomposing and linear projection of the original descriptors into a low-dimensional subspace. Detailed researches of this project includes: 1) Bound constrained optimization for the oblique image matching to relieve the influences caused by the geometric and radiometric problems of oblique images; 2) Spatial relationships constrained outlier detection to reduce the outliers in oblique image matching and improving the quality of training samples; and 3) Distance metric learning for the dimension reduction of SIFT-like descriptor, a supervised metric learning approach is proposed by exploiting the labels in the training samples of correspondences and non-correspondences. Using the semantic information originated from the labels, a linear projection is learned to embed the original descriptor to a low-dimensional space, where correspondences and non-correspondences are segregated by a large margin. Aerial oblique images collected by SWDC-5 and the ISPRS benchmark are used to evaluate the proposed methods.

航空多角度倾斜摄影测量已经成为实景精细三维模型重建的主要手段，针对倾斜影像高效处理广泛采用类SIFT高维特征描述符面临的匹配效率瓶颈难题，本项目深入研究高维特征描述符维度约减方法。传统谱分解的降维方法，如主成份分析等，依赖于数据的正态分布假设，并不适用于具有显著纹理特征的特征点。针对特征点的结构特性以及影像匹配的特殊需求，本项目拟将高维特征描述符的维度约减转化为近邻信息监督度量学习问题，不对数据分布做任何假设，仅利用训练数据中同名点与非同名点的语义标记信息，将高维特征描述符投影至低维子空间中，具体研究内容包括：1）边界约束的倾斜影像特征匹配方法，克服几何辐射质量问题，获取训练数据集；2）空间关系约束的粗差检测，提高训练数据精度；3）近邻信息监督度量学习方法，实现特征描述符维度约减，提高匹配效率。本项目将采用国产SWDC-5倾斜摄影系统和ISPRS基准测试数据进行实验分析，验证本方法有效性。

针对倾斜摄影测量中影像差异过大导致的特征匹配难题，本项目创新的提出了一种边界条件约束的最小二乘匹配方法，突破了传统最小二乘匹配方法在非线性迭代优化求解中凸显的收敛性较差的难题，将正确匹配点的收敛率提高至100%。基于上述成果，本项目进一步研究了倾斜摄影测量中的密集匹配、点云分类和三维模型优化等问题。基于本项目资助，共发表高水平SCI/EI论文11篇，项目负责人均为第一/通讯作者；授权专利1项，正在受理4项；相关成果支持倾斜摄影测量精细建模软件OSketch研发，用广泛用于我国数字城市建设，荣获省部级奖励4项，其中一等奖和二等奖各两项。完成项目既定成果要求。