Research On SAR Image Feature Classification Based On Deconvolution Network

Using microwave imaging technology,Synthetic Aperture Radar(SAR)has the characteristics of the full-day detection of long-range targets,and shows superior performance in civil and military fields nowadays.With the development of the SAR imaging technology,the resolution of SAR images is getting higher and higher.Now,how to effectively classify high-resolution SAR images has become an important research topic.This paper uses traditional classification methods and deep learning-based classification methods to achieve classification tasks separately for high-resolution SAR images.For traditional classification methods,this paper extracts pixel-level texture features and superpixel-level texture features,and then selects an appropriate classifier to classify the image.For deep learning-based classification methods,this paper chooses a deconvolution network as the base network,and then introduces respectively the more sufficient multi-scale features and the second-order covariance features of images to improve the deconvolution network.Finally,this paper develops the SAR image classification software and integrates the above algorithms,which makes it have higher application value.The main work of the paper includes the following three aspects: 1.SAR image feature classification methods based on texture features are studied.First,we extract the pixel-level Multilevel Local Pattern Histogram(MLPH)features,which are less affected by speckles.Then,we combine this feature with the SAR-SEEDS superpixel segmentation algorithm,to solve the problem of time-consuming extraction of the MLPH using a sliding window method and the operation speed is improved.Finally,a support vector machine(SVM)is used for feature classification.Besides,the classification results based on superpixel segmentation at different scales are fused to improve the classification performance.2.SAR image feature classification methods based on the deconvolution network are studied.In this paper,we use a deconvolution network for classification,which could restore the original image size layer by layer,and obtain the classification results at the pixel level.Later,we make two aspects of improvements based on the above network.First,we use the Atrous Spatial Pyramid Pooling(ASPP)module to extract the corresponding multi-scale features for each feature layer,and fuse them with the output of their corresponding deconvolution layer.The method makes full use of the multi-scale image information to improve classification accuracy.Second,we calculate the covariance matrix between the feature maps as the second-order feature of the image.In this way,the feature of the image is more fully described,the boundary of the object is more accurate in the classification result,and the classification performance is further improved.3.SAR image feature classification software is designed.The software uses the MATLAB,Python,C++ for mixed programming to achieve SAR image classification.The software mainly includes four functional modules: data reading,preprocessing,feature classification,and auxiliary functions.Among them,the feature classification module is the core function module of the software.The software modules are independent of each other,and an expansion module interface is reserved to provide convenience for users.