基于部分跨连生成对抗网络的红外舰船目标分割方法研究

Infrared ship target segmentation is an urgent problem for naval equipment system. In this proposal, we studied the problem of infrared ship target segmentation based on generative adversarial network, aiming to improve the performance of infrared ship target segmentation, and providing an effective solution for semi-supervised target segmentation. The main content includes: 1) Through collecting, cleaning, balancing and labeling, we would build an infrared ship target dataset, and we would provide baseline results for several tasks based deep neural networks to verify its practicality and validity. 2) By utilizing the region proposal network, we would obtain the region of interests of ship target; Based on this by introducing cross-connected layers and the feature selection strategies, we would establish an infrared ship target segmentation framework based on partly cross-connected generative adversarial network, selecting valid feature maps by correlational relationship. Through comprehensive analysis of different cross-connected layers and different cross-connected feature maps, we intend to provide an optimal partly cross-connected model for infrared ship target segmentation; 3) Based on the optional partly cross-connected generative adversarial network, we would build a weakly-supervised domain adaptation generative adversarial network for infrared ship target instance segmentation with limited pixel-labeled data. The achievements will not only provide an important technical supports for the application of deep learning in infrared ship target segmentation, but also improves the early warning and reconnaissance performance of modern marine weapon system in theory and application.

红外舰船目标的自动分割问题是海军装备系统迫切需要解决的问题。为提高红外舰船目标的分割性能，解决像素级标注数据有限情况下的红外舰船目标分割问题，本项目将开展基于生成对抗网络的红外舰船目标分割方法研究。研究内容包括：1)收集、整理和标注已有数据，建立红外舰船目标数据集，利用已有深度网络建立基准实验，验证其实用性；2)使用区域推荐网络，获取舰船目标的感兴趣区域；在此基础上，引入跨连和特征选择策略，建立基于部分跨连生成对抗网络的红外舰船目标分割框架，使用相关性选择跨连特征面，通过实验分析跨连方式对模型目标分割能力的影响，确定用于红外舰船目标分割的最优生成对抗模型；3)在最优生成对抗模型的基础上，构建弱监督域适应生成对抗网络，实现红外舰船目标的实例分割。本项目研究成果将对深度学习方法在红外舰船目标分割中的应用起到重要的技术支撑，对提高现代海上武器系统的预警、侦察等性能具有重要的理论意义和应用价值。

红外舰船图像具有目标小、信噪比低等特点，现有的基于全监督学习的分割算法需要大量的像素级标注提供监督信息，但获取像素级标注的成本高昂。该项目针对该不足，在构建红外舰船目标分割数据集的基础上，对红外舰船目标分割的弱监督方法进行了研究。主要研究工作包括：（1）提出了一种跨连卷积神经网络框架。使用跨连指示符表示跨连方式，分析了不同层跨连时卷积神经网络的分类和识别性能。（2）提出了一种基于弱监督和半监督学习的红外舰船目标分割方法。它首先使用一个双分支定位网络生成舰船定位图，然后使用少量的像素级标签和大量的定位图训练显著性网络，生成舰船显著图，接着将舰船显著图和图像级标签结合生成伪标签，最后利用伪标签训练分割网络并进行测试。（3）为进一步减少红外舰船目标标注工作量，提出了一种基于无监督域自适应的红外舰船目标分割方法。它首先使用图像处理方法减少可见光和红外舰船图像外观差异，然后使用空洞卷积判别网络区分源域特征和目标域特征，最后使用基于信息熵的对抗损失训练网络。总体上，该项目研究丰富了跨连卷积神经网络的设计和构造方法，扩展了红外舰船目标的分割方法。