Research On Small Sample Target Segmentation Algorithm Based On SAR Remote Sensing Images

Synthetic Aperture Radar(SAR)imaging system is a kind of radar system that uses large aperture synthesis for high-resolution imaging,which has the unique advantage of being all-weather,all-day,able to penetrate clouds,haze and dust,etc.The rapid progress of SAR technology has made the application of SAR images more and more extensive.Among them,target segmentation is one of the important areas of SAR image processing,which has high application value for many important fields such as land and resource survey,urban planning,geological disaster monitoring,environmental monitoring,geological exploration and military applications.Nowadays,with the advent of the era of big data and the improvement of computing power of computer hardware,deep learning has been developed rapidly,while great achievements have been made in the field of semantic segmentation of remote sensing images.Compared with traditional methods,deep learning-based methods can reduce the reliance on parameters and obtain more accurate results,but the presence of coherent speckle noise interference causes the segmentation of special targets and backgrounds in SAR images to become a very challenging task.Therefore,this thesis selects SAR remote sensing image dataset as the research object,and studies the target segmentation network application technology based on SAR image small sample dataset with the deep learning semantic segmentation algorithm network as the basic network architecture which is popular in recent years.For the segmentation efficiency and accuracy problems that SAR remote sensing image segmentation has been facing in recent years,this thesis proposes a solution.The main research contents of this thesis are as follows:(1)To address the problem of small sample size of the current publicly available MSTAR dataset on the web and the lack of corresponding publicly available annotated data,this thesis carries out data expansion and data tagging.Suitable special target images were first selected to form the dataset,and then the original SAR images were subjected to image pre-processing operations to make the target boundaries clearer and easier to identify.Since the imaging effect of the SAR system varies with the interference of the received wave,the MSTAR dataset is divided into two datasets: a dark pixel map and a bright pixel map.Finally,the image expansion method is used to expand the number of small sample datasets to facilitate network training later.(2)At present,the research on target segmentation techniques for SAR images mainly focuses on deep learning semantic segmentation networks such as FCN and U-Net,which have achieved good results respectively,but the network extraction speed and segmentation effect on the MSTAR dataset used in this thesis are both mediocre.In this thesis,we propose a new target segmentation network based on the Deeplabv3+ network architecture.Firstly,the feature extraction network adopts a dense connection to enhance the feature mapping relationship between modules and extract high-dimensional features;secondly,the regular convolution in ASPP is replaced with a deep separable convolution to accelerate the training efficiency of the network without affecting the segmentation accuracy,and finally,an optimised upsampling decoding Finally,an optimized up-sampling decoding structure is added to enhance the continuity of image pixels.After thorough ablation and comparison experiments on the MSTAR dataset,it is concluded that the proposed method in this thesis has better performance in SAR image target segmentation,with significant improvement in both accuracy and efficiency,and the proposed network has the best segmentation effect compared with other commonly used semantic segmentation networks.Among them,the F1 score,average intersection ratio and pixel accuracy of the proposed network in this thesis increased by 2.99%,2.91% and 3.73% respectively on the dark pixel data set,and increased by 2.73%,3.33% and 3.11% respectively on the bright pixel data set compared with the original Deeplabv3+.(3)To address the problem of misclassification of special targets and ground background in SAR image segmentation,this thesis proposes a Seg Net network with an embedded convolutional attention mechanism,which focuses more on special targets in SAR images and can segment special targets more effectively even in the case of low differentiation between targets and the surrounding environment.The proposed method outperforms several other classical semantic segmentation methods based on deep learning with advanced recognition capability and segmentation effectiveness.In particular,the F1 score,m Io U and pixel accuracy increase by 6.76%,6.48% and 6.52% respectively on the dark pixel data set,and by6.93%,6.78% and 6.93% respectively on the bright pixel data set compared to the original Seg Net network.