摘要: |
以SSD为代表的主流深度学习方法在目标检测领域取得了显著的成绩,但由于该类方法只能以矩形框给出目标的概略位置,检测结果具有很大的背景冗余区域,特别是港口密集停泊的舰船在图像中会出现区域重叠,导致误检和漏检。针对以上问题,提出了一种具有旋转不变性的舰船目标精细化检测方法,该方法综合利用可变形卷积、可变形池化、旋转的边框回归和旋转的非极大值抑制等模块的优点,借鉴MobileNet架构对网络加速,通过学习密集区域目标的几何形变,有效预测目标的旋转角度,最终以旋转的矩形框给出目标的位置。实验结果表明,该算法可实现多类舰船目标类型区分和目标朝向判定的功能,有效地解决了实际应用中的目标精确定位定向难题,提高了自动目标识别的精确性,并满足工程应用的实时性要求。 |
关键词: 舰船检测 细粒度分类 深度学习 旋转不变性 |
DOI: |
|
基金项目:国防基础科研计划(JCKY2017204B064) |
|
A Ship Target Refinement Detection Method Based on Improved SSD |
LIANG Jie,LI Lei,ZHOU Hong-li |
(Beijing Institute of Mechanical and Electrical Engineering, Beijing 100074, China;Beijing Institute of Mechanical and Electrical Engineering, Beijing 100074, China; Science and Technology on Complex System Control and Intelligent Agent Cooperation Laboratory, Beijing 100074, China) |
Abstract: |
The mainstream deep learning method represented by SSD has achieved remarkable results in the field of target detection, but since this method can only give the approximate position of the target in a rectangular frame, the detection result has great background redundant area, especially the area overlaps of ships berthed in a dense harbor will appear in the image, leading to false detection and missed detection. Aiming at the above problems, a ship targets refinement detection method with rotation invariance is proposed. The method utilizes modules such as deformable convolution, deformable pooling, rotating frame regression and rotation non-maximum suppression and uses the MobileNet architecture to accelerate the network. By learning the geometric deformation of the dense area target, this method can effectively predict the rotation angle of the target, and finally give the position of the target with a rotating rectangular frame. The experimental results show that the algorithm can realize the multi-class ship target type classification and target orientation determination function, effectively solve the target precise positioning and orientation problem in practical application, improve the accuracy of automatic target recognition, and meet the real-time requirements of engineering applications. |
Key words: Ship targets detection Fine-grained classification Deep learning Rotation invariance |