基于代价敏感的多目标识别技术研究

The ballistic target recognition, which has been an urgent and pressing mission for air and missile defense, faces the challenges of target variety, misclassification-cost asymmetry and data unbalance. Under this background, on the base of deep analysis of multi-class classification and cost-sensitive theory, this project focuses on the research on multi-target recognition based on minimum risk to decrease the misclassification risk, promote the accuracy and enhance the credibility of target recognition. The main scientific problems in this research are the design of cost-sensitive base classifier characterized by rejection, the construction of generation and transformation model for cost matrix and the decoding strategy for cost-sensitive output based on multi-classifier fusion. These key issues will be deeply studied to realize multi-target cost-sensitive classification. In order to solve the multi-target recognition problem, this project analyses the class separability based on Fisher criterion, error-correcting capability and generalization error of ECOC ensemble. On the base of the expected conclusions, this project design adaptive coding matrix based on problem data and innovate coding theory. Then, this project studies the multi-target cost-sensitive classification methods by taking the cost and priori knowledge influence into consideration to achieve effective minimum risk recognition. Moreover, the strategy of how to design rejection output fusion and decoding strategy based on cost matrix is proposed by constructing and extending performance evaluation of cost-sensitive base classifier, which can promote the stability and robustness of the algorithms. The expected results of this project can enrich the theory of multi-classification and cost-sensitive theory. Meanwhile, it can offer novel perspectives for target recognition, which is of great theoretical significance and technique value.

本项目结合弹道目标识别种类多、误识代价非对称和样本不平衡的特点，在深入研究多类分类和代价敏感学习理论的基础上，针对最小风险多目标识别技术开展研究，以降低误识风险，提高识别可靠性。通过突破具有拒识功能的代价敏感分类器的设计、代价矩阵生成转换模型的构建以及基于多分类器融合的代价敏感输出融合解码等关键问题，实现多目标代价敏感识别。项目将基于Fisher准则分析类别可分性、ECOC集成的纠错能力和泛化误差，设计基于问题域的自适应编码，解决多类目标识别问题；在综合考虑分类代价和先验知识的基础上，开展多类代价敏感分类方法研究，实现有效的最小风险多目标识别方法；通过构建和扩展代价敏感基分类器性能评估方法，解决拒识输出融合和基于代价构建融合策略的问题，提升识别算法的稳定性和鲁棒性。项目的研究成果在丰富多类分类和代价敏感相关理论的同时，为弹道目标识别研究提供新思路，具有重要的理论和技术应用价值。

本项目对弹道中段目标最小风险识别进行了深入研究，重点从ECOC多类分类、代价敏感基分类器设计以及轻量化深度神经网络构建三个方面提出了针对性的解决方案。首先从ECOC分解框架的泛化误差估计出发，分析其稳定性，为ECOC模型选择和矩阵优化提供理论支撑；提出了基于问题域的数据感知ECOC编码方法和考虑类别可分性的解码方法，在提升分类性能的同时减小编码的冗余性，为ECOC编解码最优组合提供解决策略。然后在不损失模型精度和执行效率的前提下，深入研究有效的深度神经网络模型压缩方法，发展轻量化深度神经网络，提出了基于L21范数损失函数和正则化极限学习机以及附加分类层的堆栈融合监督稀疏自编码器，从而实现目标深层次特征提取，提高目标识别的鲁棒性和稳健性，满足弹道中段多目标识别高实时性需求。最后基于代价敏感理论，提出了基于群体智能的代价敏感特征选择方法，寻找最小代价特征子集用于分类识别；分析了代价矩阵如何生成，通过引入拒绝域构造ECOC最优基分类器实现选择性分类，将经典的二值输出变为三值输出，减小误判风险；基于一维卷积神经网络，搭建了代价敏感子网络模型，提出了基于人工蚁群算法的代价敏感剪枝算法，并利用代价敏感交叉熵损失函数指导神经网络微调过程，从而实现误识别代价最小的训练目标。项目研究成果不仅可以进一步丰富代价敏感、深度学习理论，而且扩展到中段目标识别应用领域，可以进一步提高防御系统识别性能。