Research On Aircraft Detection And Recognition In SAR Images Based On Deep Learning

Synthetic Aperture Radar(SAR)is an active sensor that uses microwave for sensing.With its weather-and illumination-independent characteristics,SAR can perform all-day and all-weather observations on targets of interest which plays an important role in both military and civilian fields.Aircraft targets are important targets for battlefield reconnaissance and surveillance and are characterized by its high value and time-sensitive changes.Therefore,how to detect and recognize aircraft targets efficiently and accurately is an important issue in the field of SAR image target interpretation,and it is also one of the difficulties.At present,with the resolution of SAR sensors reaching the sub-meter level,aircraft target recognition based on high-resolution SAR images has gradually evolved from simple target detection and positioning to target recognition on specific categories and models,which traditional methods are difficult to solve.In recent years,as deep learning has been introduced into the field of SAR image target interpretation,the target detection and recognition performance has been significantly improved compared to the traditional method.Accordingly,the research on SAR image aircraft target detection and recognition technology based on deep learning method is of great significance for realizing the intelligent interpretation of aircraft targets.This paper focuses on the problem of aircraft target detection and classification recognition in large-area high-resolution SAR images,and combines the methods of deep learning to carry out systematic research.The research includes airport target detection in SAR images,aircraft target detection in SAR images and aircraft target recognition in SAR images.As a result,a complete SAR image aircraft target automatic recognition process is realized.The main work includes the following aspects:(1)As for the airport target detection in SAR images,with regard to the problems of two typical airport detection algorithms based on line segment detector(LSD)and Faster R-CNN,we propose a hierarchical detection model of airport detection in large-scene SAR images based on improved Faster R-CNN.In the proposed model,the airport candidate region is coarsely detected by the optimized long line detector at first.Then the improved Faster R-CNN remove the false alarm and make much more precise location of airport.By correcting the pooling scale and anchors size,the improved Faster R-CNN is more suitable for long linear airport target detection.On one hand,this model avoids the information loss caused by the convolution network when a large scene image is inputed.On the other,it simplifies the complicated parameter adjustment steps of traditional methods.This model realizes the fast and accurate detecion of airport target,which does good in reducing the search range for subsequent aircraft target detection and eliminating the false alarm in non-airport area.(2)In the aspect of aircraft target detection in SAR images,because the failure detection of aircraft targets with low saliency or small size is the key to limiting the detection model performance,a Retina Net network detection model fused with joint attention is proposed.By using the focal loss function of Retina Net,the network pays more attention to the hard samples and effectively solves the problem of the imbalance of the hard and easy samples in the aircraft target detection network.In addition,by introducing the joint attention mechanism of spatial attention and channel attention,the useful features in images get highlighted while the irrelevant information get suppressed,which further enhancing the Retina Net performance.Cross layer feature fusion is added to make full use of multi-scale information and improve the detection rate of small targets.Through a series of comparative experiments based on the GF-3 data,it is proved that the algorithm in this paper can realize accurate detection of different aircrafts in different complex scenes SAR image,and achieve good detection results with strong robustness.(3)With regard to the aircraft target recognition in SAR images,considering the limited amount of data,a novel algorithm based on fusion of deep feature and traditional feature is proposed.The deep convolution network can learn deep features that the traditional methods fail to extract and possesses powerful deep feature extraction function.Therefore,in this paper,we first use the deep residual network to extract the deep features of targets.Aiming at the problem that the network based on linear convolution is not rich enough to describe the features,we add the second-order response transform(SORT)to the residual network to obtain the high-order information of the target and enhance the fitting of the network to the complex nonlinear feature space.Then,based on discrete KL transform,the geometric features of the target are extracted,and the SVM classifier is trained by using both deep feature and geometric feature.The experimental results show that the algorithm in this paper outperformed the classification algorithm based on only deep feature or geometric feature,and achieves high-precision aircraft target recognition with limit data.