联合标记关系与样本特征的多标记学习研究

In many multi-label learning tasks, there often exist incomplete data problems such as lack of labeled samples, incomplete labels of identical sample, incomplete feature and so on, which directly decrease the trustworthiness of label correlation and affect the performance of multi-label classification. Toward this end, this project takes "join learning of feature and label correlation " as the clue, carrying out the following research: Firstly, because there is strong compatibility between features and labels, a multi-label learning model based on joint of geometric structure of features and label correlation is proposed; Secondly, we will propose a joint semi supervised multi-label learning model, which takes advantage of unlabeled data information to extract discriminant feature; Thirdly, explore to extract feature and label correlation based on evaluation index driving, a multi-label model based on joint evaluation-driving feature and label correlation will be proposed, which can be suitable for more application scenarios; Finally, for the problem of real value model cannot make use label correlation, we take the joint of label-special feature and label correlation into the framework of label distribution learning. The project is expected to greatly improves the performance of multi label data classification, and further provides theoretical support for the research in related fields.

在多标记学习任务中普遍存在标记样本少、同一样本标记不完整、特征缺失等数据不完整现象，直接导致标记关系可信度降低，从而影响多标记分类的性能。鉴于此，本项目以“标记关系和样本特征联合学习”为线索，开展以下方面研究：首先，针对标记关系与特征结构强关联性，研究联合标记关系与特征几何结构的多标记学习；其次，充分利用无标记数据辅助抽取鉴别特征，构建联合鉴别特征与标记关系的半监督多标记学习模型；然后，探索以评价指标为驱动反向学习特征与标记关系，研究联合标记关系与评价驱动特征的多标记学习；最后，针对实值输出模型同样存在标记强相关，研究联合标记关系与标记独有特征的标记分布学习。项目预期在多标记数据分类性能方面有显著提升，同时进一步为相关领域的研究提供理论支撑。

项目以“标记关系和样本特征联合学习”为线索，展开在遥感图像分类、植物病害分割与识别任务两方面展开相关研究，重点解决标记样本少、数据特征受限等问题，完成以下研究内容：（1）建立局部低秩邻域关系图，提出基于局部低秩表示的高光谱图像半监督分类算法；（2）引入低秩二值标记矩阵，提出基于鲁棒鉴别空谱多特征提取的高光谱图像分类算法；（3）针对CNN特征提取的盲目性，提出基于反卷积引导的植物叶部病害分割与识别模型，降低对标记样本的依赖；（4）针对图像目标大小形状不一性，提出多尺度U网络实现番茄叶部病斑分割与识别算法；（5）为提高提取特征的可解释性，提出基于RCF（ResNet50-CBAM-FCAM）网络的遥感图像场景识别算法。所提模型降低对标签样本依赖，提取的特征具有鲁棒性和可解释性，为实际应用提供理论和技术支持。