改进的深度神经网络下遥感机场区域目标检测

卫星遥感监测器下的机场区域多类目标检测在实际生活中有着重大的军用和民用意义。为了有效提升机场区域遥感图片的检测精确率，以主流目标检测方法中更快的区域卷积神经网络（Faster R-CNN）为基础框架，针对数据侧提出了ReMD数据增强算法。同时使用更具深度的残差神经网络（ResNet）以及特征融合部件-特征金字塔网络（FPN）来提取机场区域目标更鲁棒的深层区分性特征。在末端检测网络添加新的全连接层并根据目标的类间关联性组合softmax分类器以及4个logistic regression分类器进行机场区域多类目标的精确分类。实验结果表明:相比原网络改进后的网络带来了11.6%的多类平均检测精确率的提升，达到了80.5%的mAP，与其他主流网络进行对比也有更好的精确率；同时通过适当减小建议区域的输入量，可以在降低3.2%精确率的前提下将0.512 s的检测时间提速3倍，至0.173 s，根据具体任务可以合理权衡精确率和检测速度，体现了该网络的有效性以及实用性。 

Regional object detection of remote sensing airport based on improved deep neural network

The detection of multiple types of targets in the airport area under the satellite remote sensing monitor is of great military and civilian significance in real life. In order to effectively improve the detection accuracy of remote sensing images in the airport area, based on the representative deep network Faster R-CNN in the mainstream target detection method, the ReMD data enhancement algorithm is proposed for the data side. The deep ResNet network and the feature fusion component-FPN are used to extract more robust deep distinguishing features of airport area target. Finally, a new fully connected layer is added to the end detection network, and the softmax classifier and 4 logistic regression classifiers are combined to accurately classify airport area multi-class targets according to the target class correlation. Experiments show that the improvement of the original network brings a 11.6% increase in the average detection accuracy rate of the original network, reaching 80.5% mAP. Compared with other mainstream networks, it also has a better accuracy rate. At the same time, by appropriately reducing the input amount of the recommended area, under the premise of 3.2% reduction of accuracy rate, the detection time of 0.512 s is improved by 3 times to 0.173 s. According to the specific task, the accuracy and detection speed can be reasonably weighed, which reflects the effectiveness and practicability of the network.