基于深度学习的国产高分遥感影像飞机目标自动检测

Aircraft Auto-detection in Domestic High Resolution Remote Sensing Images Using Deep-learning

基于高分辨率遥感影像进行飞机目标的自动检测对精确掌握飞机的空间位置具有重要意义。由于飞机姿态不一、背景复杂、轮廓不完整等原因,导致飞机自动检测的难度较大、检测精度不高。传统飞机检测方法主要基于人工特征和机器学习分类器,在算法鲁棒性、位移、旋转不变性等方面表现欠佳。为了解决上述问题,通过引入深度神经网络模型和迁移学习策略,基于国产高分辨率遥感影像实现了飞机目标的高精度检测。具体而言,首先构建了多尺度飞机样本数据库,并基于YOLO V2目标检测模型进行迁移学习,从而提高飞机检测模型的精度和鲁棒性。以上海浦东机场GF-2影像为例进行飞机目标检测实验,实验结果表明:飞机召回率为92.25%,正确率为94.93%;通过深度学习和模型迁移可以实现小样本情况下的飞机目标高精度检测。该方法可以推广到其他地物的检测和识别中,具有较好的推广意义和价值。

It is of great significance to automatically detect aircrafts from remote sensing imagery to get their locations.However,due to aircraft posture variance,complicated background and incomplete outlines,it is challenging to achieve a high aircraft detection accuracy.Traditional aircraft detection methods are usually based on hand-crafted features and machine learning based classifiers,which is not robust enough for the translation and rotation variations.To tackle the above issues,this paper introduces deep convolutional neural network and the strategy of transfer learning to detect aircrafts from Chinses domestic satellite remote sensing images.Specifically,this paper first constructs an aircraft sample database,which consists aircrafts of different sizes and poses.Afterwards,YOLO V2 trained with natural images is utilized as the detection model and is further fine-tuned with aircraft samples to increase the robustness and performance.Experiments were done on the Shanghai Pudong airport from Chinese GF-2 remote sensing data.Experimental results showed a good performance with a recall of 92.25% and a precision of 94.93%.It is indicated that deep learning together with model transfer can get a high aircraft detection accuracy with limited training samples.The method in this paper can be generalized to other land object detection problems which shows a good promotional value.