| Synthetic aperture radar(SAR)can not be affected by climate,weather,light and other conditions,and can image the ground in all-day and all-weather.SAR has important military and civilian value.Aircraft is an important military target and vehicle,which has the characteristics of discreteness,variability and small size in SAR images.Therefore,the automatic and accurate detection of aircraft in SAR images is a research hotspot and difficulty in the field of SAR images.Currently,the automatic detection of aircraft targets in SAR images based on deep learning algorithms has replaced traditional algorithms and become the mainstream.This thesis studies the automatic detection of SAR image aircraft based on deep learning algorithm.The main research contents and contributions are as follows:(1)Aiming at the problems of low degree of automation and poor detection effect of the current SAR image aircraft detection algorithm,an efficient SAR image aircraft target automatic detection framework is proposed in this thesis.The framework uses the airport detection algorithm GCAM and the mask method to reduce false alarms in the detection results.The large-scale Gaofen-3 system SAR image experiment with 1m resolution shows that the detection effect of the framework proposed in this thesis has a significant improvement compared with the framework without airport detection and masking.(2)A Efficient Weighted Feature Fusion and Attention Network(EWFAN)is proposed in this thesis.EWFAN is based on EfficientDet-D0 framework.EWFAN introduces adaptive spatial feature fusion module,spatial attention module and CIoU loss function based on EfficientDet-D0.The adaptive spatial feature fusion module and spatial attention module can effectively fuse spatial information and improve the saliency of aircraft targets.The CIoU loss function can improve the network regression accuracy.The large-scale Gaofen-3 system SAR image experiment with 1m resolution erifies the performance of the proposed network.The average false alarm rate of EWFAN is reduced by 8.5%and the detection time are basically unaffected compared with the original EfficientDet-D0.(3)A Bidirectional Feature pyramid and Channel Shuffle Network(BFCSN)is proposed to further improve the aircraft detection accuracy.The network has three innovations:Pooling and Shuffle Module,modified weighted bi-directional Feature Pyramid Network and Area-IoU Loss function.Among them,the Pooling and Shuffle Module can well extract and fuse the features of the aircraft target;the modified weighted bi-directional feature pyramid network can promote the information exchange between the feature maps of different layers;Area-IoU Loss function can improve the regression accuracy of the network to a certain extent.Both independent test experiments and ablation experiments on Gaofen-3 data verifies the performance of the proposed network.Compared with the original EWFAN,the average false alarm rate of BFCSN is reduced by 8.48%,the average detection rate is increased by 3.69%,and the detection time is basically unaffected,which can achieve high-precision aircraft detection in SAR images. |
