Research On SAR Image Recognition Based On Target Information Enhancement

Synthetic Aperture Radar is a key sensor for many civil and military applications,and is widely mounted on a variety of platforms such as aircraft,satellites,and bombs.The continuous development of SAR imaging technology has driven the continuous expansion and deepening of SAR image interpretation research methods.This paper focuses on the refined identification of slice-level SAR targets,and research to explore the effective methods of deep learning methods in alleviating the small number of samples,detection of inaccurate positioning,and azimuth sensitivity.The main work is as follows:Firstly,in view of the insufficient number of samples for SAR images oriented to deep learning and to alleviate the usual low shape bias and high texture bias limitations of neural networks,this paper proposes a pre-processing method for three-channel pseudo-color images from the perspective of target shape information enhancement,fusing the original SAR images,Lee filtered images and segmented images of target and shadows into the network to enhance the shape bias of the network,and with the help of Soft Pool and soft pooling attention mechanisms to improve the network’s important features.With only 12.5% of the training samples of the MSTAR dataset,the recognition rate of this method still reaches 89.93%..Secondly,for the difficulty that the slicing class refinement recognition model is sensitive to the location of the anchor frame of pre-detection,a local perceptual region enhancement algorithm is proposed from the perspective of target position information enhancement to cope with different displacement cases.In this paper,we propose Sigma Pool with anti-aliasing capability,which can effectively expand the perceptual field of deep neural network for the case of small offset,and the restricted activation module is designed for the case of large target translation in combination with weakly supervised target localization to guide the network to capture the target region.The experiments demonstrate that the proposed method can expand the sensing area of the network and alleviate the degradation of recognition rate caused by target translation,and the recognition rate can still reach more than 50% when the target is translated by15 pixels..Finally,this paper proposes a general decoupling algorithm architecture that decouples the depth mixture features into two categories of eigenfeatures and variation features.For the SAR target recognition azimuth-sensitive problem,this paper designs an attention-based feature decoupling module from the perspective of target category eigeninformation enhancement,uses mutual information to optimize feature adversarial decoupling,and reduces the coupling of category eigenfeatures and azimuth-related features.The experiments demonstrates that the method in this paper reduces the correlation between features and obtains a recognition rate of 70.45% on the local azimuth data set.