鲁棒判别的多视角自适应子空间学习及其在异质图像识别上的应用研究

In order to provide a powerful guarantee of extracting features from complex multi-view images, how to find a more robust but discriminative common subspace by making full use of existent informative knowledge has been a very challenging problem in the field of multi-view subspace learning. Aiming to overcome this problem, this project combines basic theoretical disciplines from multiple different research fields and then constructs more robust and discriminative multi-view subspace learning techniques for the solution explorations for those scientific problems which have not been touched yet. To be specific, the research topics focus on four aspects, namely, 1) to develop multi-view discriminant adaptive subspace learning algorithms; 2) to steer more effective multi-view discriminant adaptive subspace learning methods using the robust cross-domain adaption, which are some extensions of the above work; 3) to propose robust maximum margin multi-view discriminant adaptive subspace learning methods based on a structurally sparse capped-norm, which are to address such a problem that the above works cannot guarantee the optimal separation and robustness in the process of feature extraction. Finally, the proposed algorithms are applied to heterogeneous image recognition for the verification of their contributions in application problems. The exact learning of multi-view subspaces in a complex environment acts as an important role in image recognition, whose relative theories enrich the theoretical system of both pattern recognition and machine learning and simultaneously have important theoretical and practical significance to promote the development of image feature extraction.

如何充分利用已有的知识,从复杂多视角图像数据中找到更具鲁棒和判别性的共享子空间,实现图像特征的有效提取已成为多视角子空间学习领域一个亟待解决的难题。针对此问题,本项目结合多个领域的相关理论方法,拟构建更具鲁棒判别力的多视角子空间学习技术探索现有研究尚未涉及科学难点的解决方案,内容包括:1)基于局部与全局鲁棒跨域适应机制的判别多视角自适应子空间学习模型;2)对上述研究进行拓展,利用鲁棒跨域适应机制驱动出更有效的判别多视角自适应子空间学习模型;3)旨在克服上述研究无法保证最优可分和特征提取过程鲁棒性问题,拟提出基于结构稀疏切范数的鲁棒大间隔多视角自适应子空间学习模型。最后,将所设计的模型应用于异质图像识别问题,考察其应用价值。复杂环境下多视角子空间精确学习在图像识别中担任着重要的角色,相关理论能够完善模式识别、机器学习理论和方法,对于推动图像特征提取发展有着重要的理论和实践意义。

本项目结合模式识别、机器学习、最优化、计算机视觉等相关理论方法，对基于结构稀疏切范数的鲁棒大间隔自适应子空间学习及其复杂图像识别问题进行了研究。具体的，研究了灵活非贪婪且具结构稀疏性的鲁棒自适应大间隔判别子空间模型；研究了非峰范数距离下的鲁棒自适应判别分析子空间学习模型；研究了结构切稀疏鲁棒主成分分析子空间学习方法；研究了鲁棒多视角和视角类间差最大化的多视角自适应图像识别模型；针对复杂多视角图像识别问题，研究了基于低秩和稀疏矩阵恢复的鲁棒子空间自适应学习模型。模型优化为本项目创新性的工作。为了便于求解，本项目从理论上给出了模型的等价形式，设计了合适有效的求解算法，理论上证明了其收敛性和有界性。为了验证所提模型的有效性，本项目在大光照变化且污染的人脸、户外交通标志、手写体、目标等图像库上与最新典型的相关算法进行了实验比较，通过计算、比较识别精度和计算代价，验证了所提方法有效性。实验显示，本项目研究的方法在识别率上，较相关流行方法有2%-10%的提高。本项目开展的相关理论能够完善模式识别、机器学习理论和方法，对于推动复杂应用环境下子空间学习技术以及复杂异质图像识别应用的发展有着重要的理论和实践意义。项目执行期内，项目组发表高水平论文8篇，申请发明专利和软件著作权4项，项目负责人得到了校杰出青年基金项目的资助，项目组成员范习健成功获批国家自然青年基金项目和江苏省双创人才，刘凡成功晋升副教授且被聘为校青年教授。