Study On SAR Imagery Detection And Recognition Based On The Fusion Of Domain Knowledge And Deep Learning

As an active microwave remote sensing imaging system,synthetic aperture radar(SAR)has the advantages of all-day,all-weather and long operation range.It has been widely studied and developed in military and civil fields.In recent years,with the rapid development of radar devices and related loads,it has become an important development trend to obtain massive wide-range and high-resolution SAR images by using airborne and spaceborne SAR technology.At the same time,how to automatically extract interesting scenes and sensitive targets from these SAR images is becoming more and more important.There are many problems in automatic interpretation of SAR images,such as automatic extraction of surface feature information,surface feature change information,target location and type information.Unlike optical images,SAR images belong to the sparse scattering center images.In addition,multiplicative noise and cross cell distribution of strong scattering points will also affect the image quality.Traditional methods automatically interpret SAR data based on the domain knowledge model in SAR image field.With the rapid development of intelligent algorithms such as machine learning and deep learning,the application of deep network to SAR image automatic interpretation has developed rapidly.However,the domain knowledge-based interpretation algorithm and deep learning algorithm face the situation that the features are independent of each other,so they are difficult to fuse with each other to improve the recognition accuracy.This paper focuses on the key problem of how to integrate the domain knowledge in the field of SAR image into the deep learning model to improve the task performance.The main contributions of this dissertation are as follows:Aiming at the water body detection problem in high-resolution SAR images in complex environment,a cascaded fully convolution network is proposed to improve the performance of water detection in SAR images.Firstly,aiming at the resolution loss caused by large stride convolution in traditional convolution neural network,a fully convolution upsampling pyramid network is proposed to suppress this loss and realize pixel level water detection.Then,considering the problem of fuzzy water boundary,the fully convolution conditional random field is introduced into the fully convolution upsampling pyramid network,which reduces the computational complexity,realizes the automatic learning of Gaussian kernel in the fully convolution conditional random field,and improves the boundary accuracy;In addition,in order to eliminate the low training efficiency caused by the unbalanced distribution of categories in the training data set,a new variable focal loss function is proposed.The detection results on Gaofen 3 SAR image show that the proposed method doesn’t only improve the pixel accuracy and boundary accuracy of water detection,but also has good detection robustness and speed.In Chapter 3,aiming at the rotation ship target detection in multi-resolution SAR images,a detection method based on anchor-free frame and key points is proposed.Firstly,the backbone network and feature pyramid network are used to extract multi-dimensional features from input SAR images,and then a sample allocation strategy of constructing head network on different level feature maps to detect ship targets with different scales is adopted.When training the network,a target center point location method based on rotating nonstandardized Gaussian function is proposed,which can correlate the orientation angle information of ship target and reduce the sensitivity of the network to target center drift.Meanwhile,aiming at the sample imbalance between multi-scale targets in the ship detection data set,a non-uniform loss weighting method is proposed to realize the effective utilization of ship targets with different scales.The detection results on the public SAR ship detection data set prove the effectiveness of the proposed method for rotating ship target detection.In Chapter 4,aiming at the limited recognition performance of SAR target recognition under different operating conditions,a fusion recognition framework integrating electromagnetic scattering characteristics and deep network is innovatively proposed,and the direct fusion based on scattering feature map and network feature map is given respectively.In the first method,the reconstructed scattering characteristic maps and the fully convolution neural network characteristic maps are fused,and the subsequent fusion processing sub-network is constructed to realize the half-way end-to-end optimization training of the model.In the second method,the scattering center vectorization is realized based on the bag of visual word model,and the effective feature analysis between feature sets is realized by discriminant correlation analysis.Experimental results show that both methods can achieve better recognition accuracy and robustness by fusing scattering center features and neural network features.In Chapter 5,aiming at the few-sample trace detection in SAR coherent change detection,the deep learning method is introduced into the coherent change algorithm,and a new trace detection paradigm is proposed,which realizes the hierarchical fusion of unsupervised coherent statistical model and supervised deep learning model.Aiming at the low correlation of difference images caused by natural factors,which seriously affects the detection performance,multi statistics based on intensity summation and intensity difference are proposed to extract water area and vegetation area,and suppress the corresponding false alarm phenomenon.Then,the coarse to fine image is constructed by using the feature classification information and trace features,and the compressed Unet is constructed to improve the utilization efficiency of trace samples.At the same time,based on unsupervised pre-training and salient transfer learning of a small number of labeled trace samples,the effective training of trace detection model is realized.The experimental results on the measured coherent change detection data show that the proposed method is effective for the detection of trace targets with few samples.In Chapter 6,according to the distribution characteristics of vehicle trace targets,Unet with spatial feature enhancement and adaptive data enhancement technology are proposed to realize vehicle trace detection.Firstly,the pseudo color image is synthesized based on the two-stage coherence estimation method to realize the effective retention of intensity change information.Then,considering the long continuity and parallel distribution of vehicle trace samples,the Unet trace detection model with spatial feature enhancement is constructed by integrating spatial CNN and spatial attention mechanism;On this basis,an adaptive enhancement strategy with registration error and multiple estimation windows is proposed to realize the effective augmentation of trace samples.Experimental results on public data and our measured data show the effectiveness of the proposed method.