摘要
由于图像数量多,因此准确、高效的目标检测是提升靶场光测图像处理自动化程度的关键步骤。针对低空目标图像及目标类型多、目标特性变化等情况导致传统目标检测算法适应性差的问题,提出了一种基于双重分类深度学习的低空目标自动检测方法。该方法基于深度学习目标检测框架YOLO V3,根据低空目标的亮度和形状的双重属性特征,将网络输出层中的单属性分类改进为双属性分类;基于目标区域生长实现样本自动标注,利用序列图像目标约束增加检测结果确认环节。靶场低空场景下的实际图像训练及检测结果表明:该方法初步检测成功率高于90%,后处理之后取得了99%的检测成功率和62%的平均定位精度。
Due to the large number of images,accurate and efficient target detection is the key step to enhance the automation of the shooting-range photometric image processing. Aiming at the problem of poor adaptability of traditional target detection algorithms due to multiple low-altitude target images and target types and changes in target characteristics, this paper proposes an automatic low-altitude target detection method based on dual-attribute classification deep learning. The method is based on YOLO V3,a deep learning target detection framework,and improves the single-attribute classification in the output layer of the network to dual-attribute classification based on the dual-attribute features of luminance and shape of the low-altitude target;achieves automatic sample annotation based on target region growth,and confirms the detection results using sequential image target constraints. The actual image training and detection results in the low-altitude scenario of the range show that the initial detection success rate of the method is higher than 90%,and 99% detection success rate and 62% average localization accuracy are achieved after post-processing.
作者
钟立军
林彬
王杰
甘叔玮
张小虎
ZHONG Lijun;LIN Bin;WANG Jie;GAN Shuwei;ZHANG Xiaohu(School of Aeronautics and Astronautics,Sun Yat-sen University,Guangzhou 510275,Guangdong,China)
出处
《上海航天（中英文）》
CSCD
2022年第2期91-98,共8页
Aerospace Shanghai（Chinese&English）
关键词
双属性分类
目标检测
深度学习
自动标注
后处理
dual-attribute classification
target detection
deep learning
automatic labeling
post processing
