Target Detection And Classification For High-resolution SAR Image Based On Contourlet CNN And Selective Attention Mechanism

Synthetic Aperture Radar(SAR)can realize all-day and all-weather observations of the earth because it is not affected by conditions such as illumination and weather,and has wide application in practice.Especially,SAR image target detection and classification have important significance in the detection and identification of military targets.The autonomous learning ability of deep learning algorithm makes it have great advantages in image classification,image segmentation,target detection and other fields,many achievements have been made in the field of SAR image processing.Based on the theory of SAR image target detection and classification and deep learning,this paper proposes a target detection algorithm and two target classification algorithms.Target detection in SAR image based on CNN and selective attention mechanism is proposed.Aiming at the problem of slow detection speed and poor detection performance in complex scenes of traditional target detection in SAR image,the proposed method introduces selective attention mechanism to rapidly calculate salient feature maps of SAR image,and then extracts target candidate region according to salient regions.Then the convolutional neural network learns the characteristics of the target candidate region and determines whether the target candidate region belongs to the target or the background,and finally obtains the final detection result through non-maximum suppression.Experiments test and verify the proposed method can quickly and accurately detect all targets in SAR image.Target Classification in SAR Image based on multi-feature SENet is proposed.Inception module in the network proposed in the method extracts the multi-scale information of the image through the different groups of convolutional filter.SE module learns and acquires the importance degree of each feature channel in network,and then improves the features that are useful to the current task and suppress the useless features according to this degree of importance,Residual module solves the problem of performance degradation caused by the increase in the number of network layers.In addition,the proposed method improves the network,which combines the low-level features of the network with high-level features to form feature information with more representative capabilities,enhancing the generalization performance of the network.By testing on the MSTAR data and comparing with the other five methods,it is proved that this method has a good classification effect.Target classification in SAR image based on NSCT_SENet and feature combination is proposed.The proposed method introduces the nonsubsampled contourlet transform into the network.After the SAR image is decomposed by NSCT,the multi-scale and multi-directional information of the SAR image is effectively extracted,and the influence of SAR image speckles and background clutter on the target is mitigated.Then the feature information is combined with the feature information of the original image extracted by the network,which makes the feature information more representative and richer.Experiments show that this method has better classification effect on MSTAR data than other methods.