Research On SAR Image Classification Algorithm Based On Convolutional Neural Network And Neighborhood Correlation

Synthetic Aperture Radar(SAR)has all-weather,all-day imaging capabilities and has been widely applied in military and civilian fields.SAR image can provide the information of the categories of terrains in SAR image.Consequently,it is one of the key step of SAR image processing.Consequently,it is important to investigate the classification of SAR images.For the SAR image classification,feature extraction has a direct impact on classification accuracy.Convolutional Neural Network(CNN)can extract deep image features that are robust and discriminative.Therefore,the CNN-based SAR image feature extraction and classification has become one of the hot topics.However,the CNN-based SAR image classification algorithm lacks the effective use of neighborhood correlation information of SAR images,which limits the classification accuracy of the algorithm.Aiming at this problem,this thesis proposes three new SAR image classification algorithms based on the combination of SAR image neighborhood correlation information and CNN deep features.The main research works of this thesis are described as follows:1)A SAR image classification algorithm using Adaptive Neighborhood-based CNN(AN-CNN)model is proposed.For CNN-based SAR image classification algorithms,the neighborhood correlations within the input image patch are not fully used,which can limit the classification accuracy of these algorithms.To solve this problem,the AN-CNN first calculates the spatial-feature bilateral distance between the neighborhood pixel and the central pixel.The feature distance-based weighting is then used to improve the classification accuracy of the boundary region and the spatial distance-based weighting is used to improve the classification accuracy of the homogeneous region.Experimental results verify the effectiveness of the algorithm.2)A SAR image classification algorithm combining SRAD(Speckle Reduction Anisotropic Diffusion)filter and CNN is proposed.Aiming at the problem that the CNN-based SAR classification algorithms cannot locate the boundaries accurately,this algorithm constructs an adaptive filtering layer using the structure of SRAD filter.The filtering layer utilizes the neighborhood correlation information contained in the input image patch to reduce the speckle noise and enhance the boundaries.The CNN layers classify the filtered image patches.By jointly adjusting the parameters of the filter layer and CNN layers,the combination of SRAD filter and CNN can be realized.Experimental results verify the effectiveness of the algorithm.3)A region-level SAR image classification algorithm which combines CNN and region-level MRF(Markov Random Field)is proposed.For the CNN-based region-level SAR image classification algorithms,the neighborhood correlations between the adjacent superpixel regions are not considered.To solve this problem,this algorithm utilizes the deep features extracted by CNN to construct the unary energy function of region-level MRF,and applies the spatial correlation constraints between adjacent regions to construct the binary energy function of RCC-MRF.Through combining the CNN deep feature within the super-pixel regions with the spatial correlation constraints between adjacent regions,region-level SAR image classification can be realized.The experimental results verify the effectiveness of the algorithm.