SAR Image Segmentation Based On Deconvolution And Adaptive Inference Network

SAR image segmentation is one of the key technologies in SAR image processing, which results directly affect the subsequent treatment, so SAR image segmentation study to become a SAR image processing inside everyone sought after topics. Combined with the current popularity of unsupervised deconvolution neural networks and classic adaptive resonance theory, proposed a SAR image segmentation based on deconvolution network and adaptive inference network,and you can get a better SAR image segmentation. The method is based on the SAR image area graph, the author gives a detailed diagram of the extraction process areas,and a SAR image include gathering regions﹑homogeneous regions and structural regions. The main work includes the following points:(1)SAR image gathering region has obvious structural characteristics,in view of this, using the proposed method-- deconvolution network and adaptive inference network to split these areas. This method first construct a multilayer depth deconvolution network, where in the first layer is the input layer, the remaining layers are deconvolution layer, in a greedy way layer by layer to take a top-down learning each layer filters and feature map. All the layers of the filter size is fixed, the greater the number of layers, the more comprehensive to learn the characteristics, contain more information, resulting in characteristic graph becomes larger, the number of filters is also increasing. Secondly, for each gathering region do not communicate with a sliding window of fixed size partition point sampled area sample, and input into the constructed depth deconvolution network to learn, the last layer of structural characteristics of the filter bank obtained by learning represent individual gathering region spatially disconnected. Finally, adaptive inference network proposed determination of similarity between the gathering areas. Take any structural feature filters of a gathering area to ART to learn, after learning to be completed, adaptive resonance theory plus regional statistical similarity computing module and rule-based reasoning module get adaptive inference network, input another gathering area’s structural features filter into the the adaptive resonance network, according to the regional statistical similarity computing module and rule-based reasoning module stars similarity between two regions, two gathering area is further divided.(2)SAR image features in homogeneous region and structure region is not obvious, not suitable deconvolution network, therefore, this paper based on the "Artificial Characteristics + machine learning" approach to its segmentation. As to SAR image of homogeneous region, extracting gray mean and variance of each region as a regional feature, then using k-means clustering method for its. As to SAR image of structural region, the paper used watershed algorithm. Finally Merge split results of each region to obtain a complete SAR image segmentation.Experiments show that the method SAR image segmentation based on deconvolution network and adaptive inference networks which this paper proposes can obtain better segmentation results.