Research On High Resolution SAR Image Classification Algorithm Based On Deep Learning

As one of the most important source of remote sensing data,Synthetic Aperture Radar(SAR)systems are irreplaceable in the field of remote sensing.High-resolution SAR images have very rich spatial structure characteristics,and it is through these characteristics that the information of the analytical image is reflected.Therefore,how to make use of the data containing rich information has become a hot and difficult point in the current SAR image research.Research on SAR image interpretation is mainly based on the feature extraction,the traditional feature extraction method is mostly artificial selection method,the applicability and robustness are not high,and the characteristics of the artificial extraction are the underlying visual characteristics,unable to involve the high-level,or cannot fully represent the essence of the target attribute,lead to poor performance of SAR image processing.The content of this paper is to analyze high-resolution SAR images based on deep learning method.The main content is as follows:(1)Aiming at the problem of SAR image scene classification,a new SAR image scene classification algorithm based on spatial feature re-calibration network is proposed.Firstly,the multi-scale spatial features of SAR images are obtained by constructing a multi-scale omnidirectional Gaussian derivative filter.Then,the detachable convolution and the additional momentum method are introduced to the characteristic re-calibration network.The network forms a bottleneck structure through the full connection layer,learns the correlation between characteristic channels and performs weight screening for multi-scale spatial features.Finally,the classification results are obtained through the training of convolutional neural network.(2)In order to solve the problems of low automation and low universality of existing high-resolution SAR image road extraction algorithm,a multi-feature extraction algorithm based on multi-path optimization network was proposed.Firstly,Gabor transform and Gray Level-Gradient Co-Occurrence Matrix(GLGCM)transform the original SAR image to obtain rich road feature information.The multi-path optimization network is formed by connecting cascade optimization network and residual network,and then the SAR original image,low-level feature image and label image obtained are trained to make full use of road features extracted from each layer network to obtain the initial road segmentation results.Finally,mathematical morphology operation is used to connect the initial road fracture and remove false alarm.the continuous innovation of deep learning accelerates the development of computer vision,effectively overcomes the limitations of existing SAR image interpretation methods,and provides a reliable method for the study of SAR image content.