基于多级宽度学习的弱小运动目标智能化并行实时检测

Dim and small objects detection is to detect true objects from complicated sequential images, which is the key point for objects tracking, recognition and the following quick reaction. It has tremendous applications in areas of national defense and military, astronomical observation, aerospace, optical remote sensing, nonmaterial, and bioscience. It is essentially detecting dim and small objects with high detection probability, low false alarm rate in real time. However, the existing detection methods are hard to achieve both the high detection performance and real-time property synchronously. This proposal starts from deep learning and big data computation in the distributed and parallel mode. The innovations are as following: (1) The parallel, multistage, multi-focus wide learning model, and learning samples generating method with image blocks combining both time and spatial features are used to estimate the complicated background of dim and small objects intelligently with high accuracy; (2) The multistage classification of dim and small objects and false alarm early rejection scheme is used to realize dim and small objects detection with high detection probability and computing resource optimization; (3) The distributed and parallel big data computation is used to design and optimize the dim and small objects detection method with high detection probability, and finally the real-time dim and small objects detection can be realized. This proposal could achieve intellectualized dim and small objects detection in real time and parallel with high detection probability, extending the application area of deep learning, and improving the research method and ability in related application area.

弱小运动目标检测是指从复杂背景序列图像中有效提取真实目标，是目标跟踪、识别及后续快速响应的关键，在国防军事、天文观测、航空航天、光学遥感、纳米材料以及生命科学等领域具有重要应用。其本质是弱小运动目标的高检测率、低虚警率、实时检测。现有方法很难同时兼顾检测性能和实时性，因此，本项目拟从深度学习和大数据分布式并行计算角度出发进行研究：（1）基于并行多级多聚焦宽度学习模型及方法、时间空间特征融合的图像块学习样本构建方法，实现弱小运动目标多复杂背景的智能化精确估计；（2）基于弱小目标多级分类及虚警早期抑制策略，实现计算资源优化的弱小运动目标高检测率检测；（3）用分布式并行计算进行弱小运动目标的智能化高检测率检测方法的并行设计和优化，实现弱小运动目标实时检测。本项目可在理论方法和技术上实现弱小运动目标智能化高检测率并行实时检测，拓展深度学习理论的应用范围，丰富和提高其在相关应用领域研究手段及能力。

弱小运动目标检测在国防军事、航空航天、光学遥感等领域具有重要应用。其本质是弱小运动目标的高检测率、低虚警率、实时检测。深度学习近年来在机器视觉、自然语言处理等诸多领域取得了巨大成功。本项目从深度学习和大数据分布式并行计算角度出发进行研究：（1）研究并行多级宽度学习网络，并进一步对模型扩展至宽度和深度神经网络模型；（2）研究基于候选区域特征学习的弱小运动目标检测方法；（3）研究基于分布式集群及GPU的并行多级宽度学习网络及弱小运动目标快速检测方法的优化实现方法。项目研究成果包括：（1）并行多级宽度神经网络模型，可实现输入影像数据及序列数据的并行、增量式处理；（2）宽度和深度学习模型，可实现输图像及非图像数据特征的高效快速提取；（3）基于候选区特征学习的弱小运动目标检测方法，可实现弱小运动目标的高检测率快速检测；（4）可扩展宽度学习模型、动态宽度和深度学习模型以及基于快速网格学习的弱小运动目标检测方法，提升并行化效率并优化计算资源。本项目的研究发展了弱小运动目标智能化高检测率并行实时检测方法，提出了并行多级宽度学习等系列增量、并行图像及非图像数据处理方法，拓展了深度学习理论及其应用范围。本项目发表SCI论文10篇，EI论文6篇，申请发明专利8项，多篇论文发表于领域内高质量期刊上。