Research On SAR Image Target Detection Technology Based On Deep Learning

With the development of Synthetic Aperture Radar(SAR)and the improvement of SAR imaging technology,the obtained SAR images have more information and are widely used in civil and military fields.Therefore,the interpretation of SAR images is required.It also brings more challenges.SAR image target detection plays a very important role in the process of SAR image interpretation,and it is a heat research in the field of SAR image processing.Deep learning has developed rapidly in recent years and has shown excellent characteristics in target detection.Thesis studies the target detection technology of SAR images based on deep learning.Firstly,the main features such as geometric features and intensity features of SAR images are analyzed.The classic statistical model based on Constant False Alarm Rate(CFAR)SAR image target detection algorithm is studied.The adaptive detection under different statistical models is analyzed.The threshold process is introduced on deep learning theory.Then,the machine learning algorithm is studied from the perspective of supervised learning and unsupervised learning.Logistic regression and support vector machines are mainly used for the two-classification problem.Logistic regression and SVM classifiers are extended to multi-classification algorithms,and the SIFT features of the target are used.The training of the classifier completes the classification task of the SAR image target.Finally,the realization of SAR image target detection network based on convolutional neural network is studied.A convolutional neural network is constructed based on SAR image target and SAR image dataset characteristics.A convolutional feature fusion network is proposed,which combines multi-level convolution features and performs Local Response Normalization.The feature enhances the feature expression ability of the small target of SAR image in complex environment.The target network is combined with the SVM classifier in machine learning to obtain the complete network used in this paper,and the target detection is obtained and the target category information is obtained.In the experiment,based on the relatively complete SAR target slice data,the SAR image expansion training samples in complex large scenes are added,and the whole network is optimized through end-to-end approximate joint training to realize the whole SAR image detection of the target in different complex large scenes.