基于视觉注意机制的SAR图像小目标检测方法研究

Automatic detection of man-made targets such as ships, planes and cars in SAR imagery is of much significance both in military and civil applications. However, these targets usually occupy a small portion of the image and lack of detailed feature information. Therefore, the existing methods suffer from inefficiency problems, characterized by a slow computation speed and a lack of accuracy when the image is large or the background is complex and cluttered. In this project, computational models of visual attention are introduced for interpreting SAR images and a new target detection method based on the visual attention mechanism is proposed for automatic detection of small targets in SAR images. According to features of SAR images and characteristics of small targets, this proposed project will be composed of three parts. The first part will focus on designing a SAR-suited visual attention model that solves challenges in feature extraction and selection, scale selection, saliency map generation, as well as focus selection and changing. The second part will focus on improving the proposed visual attention model by learning from the advantages of the spectral analysis models. And the last part will focus on using contextual information to refine target detection. Specifically, the description methods of contextual information and the control strategies of top-down attention are studied. A target detection scheme based on double-attention will be designed to use the bottom-up saliency map and top-down contextual information, therefore enabling accruate detection of small targets under complicated backgrounds. This proposed project will enhance the development of intelligent processing of SAR image and provide technical supports for tasks such as military target searching and monitoring, marine target detection, and intelligent transportation.

对SAR图像中船只、飞机、汽车等人造目标进行自动检测在军事和民用领域均有重要的应用价值。然而,这类目标在图像中有效能量区域小且细节信息比较匮乏，现有方法在检测精度和效率方面均存在不足。本项目旨在将视觉注意的模型化计算引入SAR图像解译中，基于视觉注意机制构建一套高效稳健的小目标检测方案，为SAR图像目标检测研究提供新思路。根据SAR图像特征和小目标特点，研究模型化计算过程中涉及的特征选取、尺度选择、显著图生成、注意焦点选择等问题，设计出适用于SAR图像的视觉注意计算模型；针对大视场、强杂波、点目标等情况，借鉴变换域分析方法的优势提出改进措施；引入目标所处的上下文环境进行自顶向下的控制，研究上下文信息的描述方法及控制策略，构建融合双向注意机制的目标检测方法，实现复杂背景小目标检测。本研究有望推动SAR图像智能化信息处理的发展，并为军事目标搜索与监视、海上目标检测、智能交通等提供技术支持。

随着SAR图像分辨率提高，传统目标检测方法往往存在统计模型参数估计复杂、对大视场图像检测精度与效率难以兼顾等问题。特别是对于一些小型人工目标如车辆、船只、飞机等，由于目标自身不变特提取困难，难以满足实用性要求。课题将视觉注意的模型化计算引入SAR图像解译中，围绕SAR图像特征提取与优选、适用于SAR图像的视觉注意计算模型、基于视觉显著性的目标检测方法等展开了深入研究。通过三年的研究，取得了如下成果：（1）针对SAR图像特点，以特征整合理论为基础，构建了基于多特征融合的视觉注意模型，并拓展到全极化SAR图像；（2）针对强杂波背景和小目标，改进了频率域视觉注意模型；（3）模拟人类视觉过程，设计了基于视觉反差和信息熵的SAR图像目标检测方法，并利用上下文线索进行约束，提出了融合双向注意机制的目标检测算法，应用于SAR图像舰船目标和车辆目标检测中；（4）设计了基于混合编码遗传算法的SAR图像特征优选方法；（5）提出了基于颜色恒常性和视觉注意的阴影检测方法；（6）开发了SAR图像处理与人工目标检测原型系统，建成典型目标样本库和不同场景测试图集，并开展了不同模型性能评测研究。项目研究为SAR图像目标检测提供了新思路，能为军事目标搜索与监视、海上目标检测、智能交通等在目标预检测阶段提供技术支持。在项目资助下，发表学术论文9篇，申请国家发明专利4项，培养博士研究生1名、硕士研究生7名。