多任务学习在SAR图像目标识别中的应用

发布时间：2018-10-19 07:14

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)是一种全天时、全天候、主动对地观测传感器,实现SAR图像目标识别具有重要意义。由于SAR图像获取成本高以及SAR图像目标姿态敏感性,导致用于目标识别的带标签SAR图像样本不完备,对SAR图像目标识别带来挑战。多任务学习(Multi-task Learning,MTL)利用不同信息源或特征,同时学习多个回归模型优化参数,实现多特征信息融合,有利于提高识别性能。本文基于SAR图像多尺度特征的稀疏表示,研究MTL架构中的特征选择策略、稀疏表示、稀疏求解等问题,主要完成工作如下:(1)为满足MTL对多尺度特征在稀疏域中空间分布相似性的要求,提出一种基于稀疏向量分布相似度的特征选择方法。首先,对验证集样本进行多尺度特征稀疏表示,在不同尺度下,按类别统计稀疏度分布,定义尺度间的稀疏度分布相似度矩阵,求得对应的相关信息熵。最后,选择相关信息熵最大的特征子集。通过实验分析特征的冗余性和目标识别率,验证了特征选择方法的有效性。(2)针对训练样本量不充足时,稀疏表示自由度偏高,提出一种多尺度特征的局部线性约束稀疏字典优化方法。基于MTL的架构,建立多尺度特征局部线性约束,降低稀疏表示自由度,实现稀疏字典的优化,提高了样本不充足下的目标识别率。实验表明,在训练样本不充足时,与联合稀疏表示相比,本文方法提升了目标识别效果。(3)设计了一种多尺度邻域加权的匹配追踪算法。在MTL的架构下,通过对残差的多尺度稀疏向量进行邻域加权,选择原子,实现匹配追踪,得到多尺度联合稀疏系数。在不同尺度下按类别稀疏重构,依据多尺度累加重构偏差,实现目标分类。实验结果表明该算法的重构精度与凸优化方法相当并且耗时较短。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture Radar,SAR) is a kind of all-day, all-weather, active earth observation sensor to realize SAR image target recognition. Because of the high cost of SAR image acquisition and the sensitivity of target attitude in SAR image, the labeled SAR image samples for target recognition are not complete, which brings challenges to target recognition in SAR image. Multi-task learning (Multi-task Learning,MTL) utilizes different information sources or features and simultaneously learns multiple regression models to optimize parameters so as to achieve multi-feature information fusion which can improve the recognition performance. Based on the sparse representation of multi-scale features of SAR images, this paper studies the feature selection strategy, sparse representation and sparse solution in MTL architecture. The main contributions are as follows: (1) in order to meet the requirements of MTL for spatial similarity of multi-scale features in sparse domain, a feature selection method based on sparse vector distribution similarity is proposed. Firstly, the multi-scale feature sparse representation of the validation set samples is performed. According to the different scales, the sparse degree distribution is calculated according to the category, and the similarity matrix of the sparse degree distribution between scales is defined, and the corresponding information entropy is obtained. Finally, the feature subset with the largest entropy is selected. The effectiveness of the feature selection method is verified by analyzing the redundancy of the feature and the target recognition rate. (2) the sparse representation degree of freedom is high when the training sample is not sufficient. A local linear constrained sparse dictionary optimization method with multi-scale features is proposed. Based on the framework of MTL, the local linear constraints of multi-scale features are established, the degree of freedom of sparse representation is reduced, the sparse dictionary is optimized, and the target recognition rate under insufficient samples is improved. Experiments show that the proposed method improves the target recognition performance compared with the joint sparse representation when the training samples are not sufficient. (3) A multi-scale neighborhood weighted matching tracking algorithm is designed. In the framework of MTL, the multiscale joint sparse coefficients are obtained by neighborhood weighting of the multi-scale sparse vectors of the residuals and the selection of atoms to achieve matching tracing. According to the multi-scale cumulative reconstruction deviation, the target classification can be realized by sparse reconstruction according to different scales. The experimental results show that the reconstruction accuracy of the algorithm is quite similar to that of the convex optimization method and the time is short.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN957.52

【参考文献】