An Aircraft Target Detection Method Based on Regional Convolutional Neural Network for Remote Sensing Images

The aircraft target is the key object of battlefield surveillance and reconnaissance. It can accurately and efficiently detect the aircraft target from the remote sensing images which aim at ground reconnaissance. First of all, it can quickly acquire the intelligence of the enemy's military activities and provide support for the identification of the air target. Second, it can evaluate the importance of the military airport and analyze enemy's operational intentions, to achieve a precise strike against the enemy's aircraft targets. The existing aircraft detection method uses a single convolutional neural network to accomplish the whole process of feature extraction and recognition. It fails to effectively extract the characteristics of aircraft targets and ignore the scale differences of different aircraft. Thus, the recognition results are not accurate enough Aiming at this problem, this paper uses the deep residual network to extract the characteristics of aircraft targets, studies and analyzes the size of different aircraft targets, and uses K-means to cluster different sizes. The cluster centers are representative aircraft sizes. Based on these representative sizes of the aircrafts, the Aircraft Targets Region Proposal Network (ATRPN) is proposed to synthesize the geometric characteristics of different aircraft. Based on the faster regional convolutional neural network detection framework (Faster R-CNN), taking the deep residual network and ARPPN as the front end and the candidate box generation network, the ATRPN R-CNN remote sensing image aircraft target detection method is proposed. This paper also establishes an aircraft target detection data set with uniform distribution, complete shape and rich aerial photography angle. After training the ATRPN R-CNN remote sensing image aircraft target detection method on the data set, the performance comparison experiment was carried out with the detection framework of Faster R-CNN and single network target multi-scale detection framework (SSD). The experimental results show that the detection method has higher detection accuracy in many different scenes including different aircraft targets. (Abstract)