Study On The Methods Of SAR Image Target Detection Based On Convolutional Neural Networks

Synthetic Aperture Radar(SAR)images target detection occupied an important role in SAR image interpretation,and with the continuous development of SAR imaging technology,target detection has become the momentous research direction in many fields such as battlefield reconnaissance and geoscience remote sensing.As the capability of SAR data acquisition has got increasingly mature,it entails the technology of SAR image processing and interpretation with high efficiency and accuracy.Benefited from the rapid development of Deep learning methods applied in computer vision field,the limitation of time-consuming and poor adaptability existing in traditional methods can be offset by the fitting capacity of deep neural networks.Aiming at laying technical foundation for future military reconnaissance and rapid intelligence search,this paper researches into the technology of SAR image target detection based on Convolutional Neural Networks(CNN),and mainly focuses on the property of small size and densely packed targets.By analyzing the challenges of deep learning methods applied to SAR target detection and summarizing the current research status and application,this paper modifies the region proposal based methods and the bounding box regression based methods respectively.The main contents of paper are as follows:1.On the basis of the modeling process of neural network,the core components of CNN are described in detail from the perspectives of convolution layer,pooling layer and Softmax classifier,and the back propagation algorithm is deduced.Then,the implementation process of Faster RCNN and SSD of two typical detectors is further studied,and the bounding box regression mechanism and default box matching strategies of two typical detectors are elaborately analyzed.2.On account of the missing detection of small scaled and densely arranged targets as well as a large amount of false alarms appearing in complicated large scene SAR images,the thought of dense feature fusion are utilized to increase the diversity of the extracted candidate region.Meanwhile,combined with the context information assisted small target positioning,a novel two-stage SAR ship detection method is formulated.In order to fully train the difficult samples,this paper also improves the loss function of the training process and solves the issue of unbalanced positive and negative samples and indistinguishable similar objects by selecting proper parameters through extensive experiments,which indicate the superiority of refined two-stage methods.3.For the purpose of meeting the real-time demand in practical application,we modify the one-stage based methods with low accuracy but high speed.First,the attention mechanism is applied to enhance the feature extraction capability of the network,and then considering the fact of enriching semantics in detected feature maps,a one-stage method based on feature utilization and enriched semantics is proposed.This method introduces the segmentation module and the two-level attention mechanism to enhance the semantic information of low-level and high-level features respectively,and the adaptability of the method is verified on NWPU VHR-10 dataset.4.A one-stage detector based on the regression of coarse-to-fine two-stage method was designed,which integrates merits including the accuracy of two-stage method with the high speed of one-stage detector.The anchor refined two-stage regression is realized by reverse connection and the model is trained end-to-end with multi-task loss function.The experimental results shows the proposed method can greatly improve the detection performance of small targets and has good generalization ability.