基于特征学习与迁移的复杂环境下SAR车辆目标检测与识别

Due to the complexity of the scene, the variation of the target, the hardness to obtain the training samples, and the insufficiency of image information exploration, the following problems still exist in the synthetic aperture radar (SAR) vehicle target detection and recognition under complex environment: man-made clutter false alarms are high for the detection and the recognition performance is not good when there are only a small number of training samples. Aiming at these problems, this project will study the following contents based on the theory of feature learning and transfer. 1) This project will study the target detection method based on the fusion of multi-scale superpixel-level multiple visual features. By proposing the detection method based on the feature fusion using one-class classification with multiple kernels, this project will make good use of the abundant visual features in the high-resolution images to improve the target detection performance.2) Considering the characteristics of SAR target discrimination, this project will construct the semi-supervised common and category-specific visual dictionary learning method aiming at the imbalanced data, to complete the SAR image content description. Then this project will complete the vehicle target and man-made clutter discrimination based on the image content. 3) When the training samples in the full aspect angle range are available, this project will study the recognition methods based on the non-linear multiple feature learning, in order to automatically learn the multiple group of features that are most effective to the classification, thus to improve the recognition accuracy. When the training samples in the full aspect angle range are not available, this project will study the recognition methods based on the feature-representation-transfer, in order to transfer the feature representation knowledge learned in the cooperative target recognition to the non-cooperative target recognition with a small number of training samples, thus to improve the recognition performance when there are only a small number of training samples. The project will finally improve the SAR target information extraction ability.

由于场景复杂、目标多变、训练样本难获取、图像信息挖掘不充分，复杂环境下的SAR车辆目标检测与识别仍面临检测中人造杂波虚警高，小样本下识别性能差的问题。针对该问题，本项目将基于特征的学习与迁移开展以下研究：1）研究基于多尺度超像素级多层视觉特征融合的目标检测方法，提出基于多核一类分类的特征融合检测算法，充分利用高分辨率图像中丰富的视觉特征提高检测性能；2）结合SAR目标鉴别特点,建立针对非平衡数据的半监督共享与类别特定视觉词典学习模型，完成SAR图像内容表示，实现基于图像内容的目标与复杂人造杂波鉴别；3）当训练样本方位角完备时，研究基于非线性多特征学习的识别方法，自动学出对分类最有效的多组特征，提高识别精度；当方位角不完备时，研究基于特征表示迁移的识别方法，以将合作目标识别中学到的特征表示知识迁移到小样本下非合作目标识别中，提高小样本下识别性能。最终提高复杂环境下的SAR目标信息获取能力。

由于场景复杂、目标多变、训练样本难获取、图像信息挖掘不充分，复杂场景下的SAR车辆目标检测与识别面临着检测中人造杂波虚警高，小样本下识别性能差的问题。针对该问题，本项目开展了基于特征表示学习与迁移技术的SAR目标检测与识别。主要研究内容和成果包括以下几个方面：1）研究了基于多尺度超像素级多层视觉特征融合的目标检测方法，包括基于回归核的散射机制差异描述方法以及基于该方法的极化SAR目标检测方法、基于超像素级多种局部信息度量融合的极化SAR目标检测方法；2）研究了基于图像内容的目标与复杂人造杂波鉴别，包括基于多特征融合词包模型的SAR目标鉴别方法、基于词包模型以及共享与类别特定字典学习的SAR目标鉴别方法、基于多卷积神经网络特征融合的SAR目标鉴别方法以及基于半监督卷积神经网络与特征约束的SAR目标鉴别方法；3）研究了基于非线性多特征学习的SAR目标识别方法，包括带有相关性约束的结构化核字典学习方法与基于密集卷积神经网络与Fisher判别正则化度量学习的舰船分类方法；研究了基于特征表示迁移的小样本下目标识别方法，包括基于角度旋转生成网络的小样本下SAR目标识别方法。本项目的研究成果可为复杂场景下的SAR目标信息获取提供技术支撑。