基于最近特征线的图像特征提取算法分析-analysis of image feature extraction algorithm based on nearest feature line.docxVIP

若 干 子 方 向，然 后 利 用 二 维 线 性 判 别 分 析 对 每 个 子 方 向 降 维，此 外，采 用 自 适应的方法计算每个子方向的判别力，增强算法的鲁棒性。4. 针对二维线性判别分析和邻域判别最近特征线分析直接无法应用在单 样 本 问 题 上 的 缺 陷，提 出 了 利 用 轮 廓 波 变 换 构 造 新 训 练 样 本 的 算 法，为 单 样 本问题的解决提供新的思路。关 键 词： 特 征 提 取；子 空 间 学 习；最 近 特 征 线；轮 廓 波 变 换；离 散 分 数 阶 余 弦变换AbstractImage Recognition is applied in many areas and has become a hot research topic recent years. Image feature extraction is a key step of image recognition. In statisti- cal pattern recognition, generally, the main idea of feature extraction is to project the original samples data to a low-dimension feature space, in which the samples should be distinguished easily. Most existing feature extraction algorithms are based on Euclidean distance. Because of the complexity of the distribution of image samples data, Euclidean distance may be not the optimal when it is used to measure the scatter of the image sam- ples. Nearest feature line (NFL) is a feature line measure based classifier, which has good performance in image classification and retrieval. However, it is only a classifier. In or- der to improve the learning ability, novel NFL based subspace learning algorithms are needed. This dissertation focusses on the NFL based image feature extraction algorithms and try to solve the high computational complexity and the unstablility of the NFL based algorithms to improve the feature extraction. The main contributions are as follows:For the problems of the existing NFL based subspace learning algorithms, the dissertation proposed a Simplified Nflspace algorithm and a Neighborhood Discriminant Nearest Feature Line Analysis (NDNFLA). The Simplified Nflspace only minimizes the distance between the sample and the corresponding feature line constructed by its two nearest neighbors in the same class, rather than the distance between the sample and all feature lines in the same class, so the computational complexity is reduced. NDNFLA aims to minimize the distance between sample and the feature lines constructed by its seve