Aircraft Target Detection Technology For Complex Remote Sensing Images In Airport Areas

In recent years,target detection technology for optical remote sensing image has rapidly developed and gradually become an important research direction in the field of image processing.As an important branch of it,automatic detection technology for aircraft targets,it plays an important role in both military and civilian fields.Although significant progress has been made in current detection algorithms,there are still challenges for aircraft target detection tasks in remote sensing images of airport areas.Firstly,the complex scene of airport areas with numerous artificial structures results in a large number of interference targets that are similar to the target features,making it difficult to distinguish between the background and the aircraft target,which easily leads to false alarms.Secondly,there are various types of aircraft in the same optical remote sensing image,with significant differences in scale,making traditional deep learning methods prone to false positives and missed detections.Finally,aircraft targets in airport areas often exist in densely parked situations,where horizontal frame detection methods are prone to missed detections and have poor visual detection effects.Current mainstream deep learning target detection algorithms have not made targeted improvements for the above challenges.Therefore,based on the theoretical knowledge of deep learning image processing,this paper proposed an aircraft target detection technology based on the Efficient Net backbone network for complex remote sensing images of airport areas.The main research work and achievements of this paper are as follows:Firstly,this paper constructed a feature extraction backbone network based on the Global Attention Mechanism(GAM)to address the challenge of complex scenes in airport areas in optical remote sensing images.The GAM attention mechanism is utilized to enhance the feature of airplane targets,thereby improving the ability of the backbone network to extract features of airplane targets.At the same time,the shallow convolution layers of the backbone network are optimized to improve the efficiency of extracting features of airplane targets in complex scenes.Secondly,in response to the challenge of diverse aircraft types,significant differences in scale,and the difficulty of mis-detection and omission,this paper proposed an Adaptively Spatial Feature Fusion(ASFF)module based on Recursive Feature Pyramid(RFP)on top of a GAM-based attention mechanism backbone network.By utilizing the useful aircraft feature information multiple times through the RFP pyramid,the network’s feature representation ability is improved.Then,the ASFF feature fusion module is connected via a fully connected method to address the inconsistent feature issue in the pyramid structure,helping the network to fully integrate feature information of different scales,thus solving the challenge of large scale differences in different types of aircraft.Finally,to address the challenge of dense parking of aircraft targets in airport areas,this paper adopted a directional bounding box labeling and detection method,constructing a Pixels-Io U loss(PIo U loss)function to calculate the regression loss of directional bounding boxes.By introducing a rotation angle parameter and optimizing the global Intersection over Union(Io U)in a pixel-wise manner,the regression of directional bounding boxes is more accurate,thus solving the problem of dense parking of aircraft targets.In summary,this paper proposed an aircraft target detection technology for complex remote sensing images in airport areas.Qualitative and quantitative experiments were conducted on a subset of aircraft in the large public UCAS＿AOD dataset.The experimental results verify the effectiveness of the proposed method for aircraft target detection in optical remote sensing images of airport areas.