中文摘要：

针对军用机场大尺寸卫星图像中航空器检测识别的具体应用场景,建立了一套实时目标检测识别框架,将深度卷积神经网络应用到大尺寸图像中的航空器目标检测与识别任务中。首先,将目标检测的任务看成空间上独立的bounding-box的回归问题,用一个24层卷积神经网络模型来完成bounding-box的预测;然后,利用图像分类网络来完成目标切片的分类任务。大尺寸图像上的传统目标检测识别算法通常在时间效率上很难突破,而基于卷积神经网络的航空器目标检测识别算法充分利用了计算硬件的优势,大大缩短了任务耗时。在符合应用场景的自采数据集上进行测试,所提算法目标检测实时性达到平均每张5.765 s,在召回率65.1%的工作点上达到了79.2%的精确率,分类网络的实时性达到平均每张0.972 s,Top-1错误率为13%。所提框架在军用机场大尺寸卫星图像中航空器检测识别的具体应用问题上提出了新的解决思路,同时保证了实时性和算法精度。

英文摘要：

Aiming at the specific application scenario of aircraft detection in large-scale satellite images of military airports, a real-time target detection and recognition framework was proposed. The deep Convolutional Neural Network （CNN） was applied to the target detection task and recognition task of aircraft in large-scale satellite images. Firstly, the task of aircraft detection was regarded as a regression problem of the spatially independent bounding-box, and a 24-layer convolutional neural network model was used to complete the bounding-box prediction. Then, an image classification network was used to complete the classification task of the target slices. The traditional target detection and recognition algorithm on large-scale images is usually difficult to make a breakthrough in time efficiency. The proposed target detection and recognition framework of aircraft based on CNN makes full use of the advantages of computing hardware greatly and shortens the executing time. The proposed framework was tested on a self-collected data set consistent with application scenarios. The average time of the proposed framework is 5. 765 s for processing each input image, meanwhile, the precision is 79.2% at the operating point with the recall of 65.1%. The average time of the classification network is 0.972 s for each image and the Top-1 error rate is 13%. The proposed framework provides a new solution for application problem of aircraft detection in large-scale satellite images of military airports with relatively high efficiency and precision.