基于大样本的海上复杂环境下弱小目标检测研究

Target detection method based on depth learning is one of the key research directions in the field of target detection, which can effectively distinguish between background and prospects, the target changes and robust to block. The existing depth learning detection model depends on the quality of training set and training methods and cannot achieve online adaptive update. However, there are few studies on the pre-training depth model considering the quality of large sample data sets..Due to our Nansha and Xisha Island need to carry out the research on small and weak target detection in complex scene at sea, this project intend to expand the existing large sample data set training ideas and propose a dynamic training model considering the quality of samples from the adaptive point of view. At the same time, this project intend to bring the idea of training the weight of the sample into the large sample training depth model. These include: considering the pre-training depth model of the quality of large sample data sets, considering the adaptive marine visual target representation and model renewal research, considering the regression method of deep learning real-time sea target detection research, and real-time detection of small and weak targets on the sea. The research idea of this project has certain innovation and engineering practicability. The expected results can be directly applied to the field of large-scale continuous dynamic monitoring of unmanned aerial vehicles.

基于深度学习的目标检测方法可有效区分背景和前景、对目标变化和遮挡更加鲁棒，是目前目标检测领域的热点研究方向之一。现有深度学习的目标检测模型在很大程度上依赖于训练集质量及训练方法，但对于考虑大样本数据集质量的预训练深度模型的研究较少，同时不能实现模型在线自适应更新。基于此，本项目以我南沙和西沙岛屿亟需开展海上复杂场景下弱小目标检测研究为背景，拟扩展现有大样本数据集训练思路，从自适应角度出发推导一种考虑样本质量的动态训练模型；同时将训练样本权重的思想引入大样本训练深度模型中，设计相应的逐个样本动态权重模型。具体包括：考虑大样本数据集质量的预训练深度模型研究、考虑自适应的海上视觉目标表述与模型更新研究、考虑回归方法的深度学习实时海上目标检测研究，最终实海上弱小目标实时检测目的。本项目研究思路具有一定的创新性和工程实用性，其预期成果可直接应用于基于无人机的海上大范围连续动态监控领域。

为了提高海上复杂环境下弱小目标检测的实时性和精度，实现基于无监督的海上大范围连续动态实时监控，本研究就基于大样本的海上复杂环境下弱小目标检测关键技术出发，扩展了现有基于大样本数据集的训练思路，对基于大样本数据集质量的预训练深度模型、基于自适应的海上视觉目标表述与模型更新、融合回归方法的深度学习实时海上目标检测等多个方面展开了研究。为了总体上更精确的表达和检测海上不规则小目标，尤其是船舶检测，还对旋转矩形框的方法、YOLO v5算法、残缺旋转矩形框优化方法、外接水平矩形框IOU等方法展开研究；同时，在研究过程中发现海天背景下的舰船类目标特征具有单一性和弱形变性，为了充分提取这类目标特征，提高算法实时性，展开了适应目标几何形状的卷积核网络、图像去燥和背景建模、分层检测海陆图像分割等方法展开了研究；展开了融合目标全局和局部表示模型、融合目标模型和遮挡模型的目标表示、目标对角外观模型和结构化局部稀疏表示等方法的研究；最后，为实现无人设备的海上大范围连续动态实时监控的目的，对海上复杂环境下的单目标和多目标的跟踪方法展开了研究。本项目研究成果对于提高海上交通监控能力、海上高精度信息快速感知、海上大范围场景的连续动态实时监控具有重要的理论意义和实际应用价值。