面向低质量图像数据的低秩判别迁移子空间特征抽取研究

With the development of information technology and the ongoing emergence of new domains, the image data acquired by people is fiercely increasing in terms of dimensionality and cannot also be guaranteed to be independent and identically distributed. How to extract the features from these image data is a very challenging problem. There are few efforts on transfer subspace learning that have made some preliminary attempts to solve this problem. However, still there exist many unsolved problems. Based on the theory of low-rank transfer representation, this project will design some novel low-rank discriminant transfer subspace feature extraction algirhtms with high discrimination for images.To be specific, the reseach topics of this project maily focus on four aspects, namely, 1)to propose a low-rank discriminant transfer subspace feature extraction algorithm via the robust distribution estimation, which considers both advantages and disadvantages of the low-rank transfer representation and the data distribution estimation stimutaneously;2) to design a low-rank preserving discriminant transfer subspace feature extraction algorithm, an extension of the above work, which aims to mine and utilize the underlying discriminant knowledge inside the low-rank transfer representation; 3) to develop a maximum slack margin low-rank discriminant transfer subspace feature extraction algorithm, which is to solve the problem that the above works cannot absolutely guarantee the extraction of the optimal discriminant features for testing image data.Finally, these algorithms are applied on low-quality image recognition problems to verify their effectiveness. The project development will give the research on transfer subspace learning a shot in the arm and simultaneously enrich the theoretical system of feature extraction.

随着信息技术的发展和新领域的不断涌现，人们获取的图像数据不仅维数上迅猛增长，分布上也难以保证满足独立同分布原则。如何从这样的数据中提取有效特征是一个极具挑战的问题。有关迁移子空间特征抽取研究在此方面做了初步的尝试，但仍有很多问题尚未解决。本项目以低秩迁移描述理论为基础，拟设计更具判别力的低秩判别迁移子空间特征提取算法。研究内容包括:1)平衡低秩迁移描述和分布估计的优劣，拟提出基于鲁棒分布估计的低秩判别迁移子空间特征提取算法；2)充分利用低秩迁移描述中潜在判别知识，对上述研究进行拓展，拟提出低秩保留判别迁移子空间特征抽取算法；3)针对上述研究不能保证测试集上判别特征最优提取问题，拟设计最大松弛间隔低秩判别迁移子空间特征抽取算法。最后，在低质量图像识别问题上，对所提算法有效性进行验证和定量分析。本项目的开展必将为迁移子空间分析研究注入新鲜的血液，丰富和发展模式识别、图像特征提取技术的理论体系。

随着信息技术的发展和新领域的不断涌现，人们获取的图像数据不仅维数上迅猛增长，分布上也难以保证满足独立同分布原则。如何从这样的数据中提取有效特征是一个极具挑战的问题。有关迁移子空间特征抽取研究在此方面做了初步的尝试，但仍有很多问题尚未解决。本项目以低秩迁移描述理论为基础，设计了更具判别力的低秩判别迁移子空间特征提取算法。具体的，从理论层面给出了基于图嵌入回归子空间特征抽取中经典的随机游走SRW方法的正则化框架，研究了其与已有的标记繁衍方法关系，提出了迭代正交化的基于图嵌入回归迁移子空间特征抽取模型，解决了判别投影不相关和投影向量数量受限问题；从理论上得出了已有的局部判别子空间学习技术本质上是通过加权实现图像数据局部拓扑结构的挖掘，但无法保证图像的鲁棒迁移能力，为此，研究了基于L1/Lp范数最大最小距离的鲁棒判别子空间特征抽取和聚类模型；研究了快速正交鲁棒判别迁移子空间特征抽取模型，其创新点在于应用了QR和回归联合模型加速了模型的计算，模型求解最终简单归结为一个回归问题；揭示了已有的鲁棒判别子空间特征抽取方法的内在联系，提出了非峰鲁棒判别模型。最后，在受光照变化、腐蚀、遮挡大的人脸、地形、手写体等图像集上，通过计算、比较识别精度和计算代价，验证了所提算法的有效性。本项目的开展能为迁移子空间分析研究注入新鲜的血液，丰富和发展模式识别、图像特征提取技术的理论体系。