SAR Image Segmentation Based On Semantic And Ridgelet Deconvolutional Network

SAR image segmentation is the foundation of SAR image processing. The results of the segmentation have a directly influence on the subsequent analysis and identify. Usually, the conventional SAR image segmentation methods are mainly divided into the method based on the characteristics and the method based of statistical model. In the method based on characteristics, the characteristics of good or bad often become the bottleneck of the SAR image processing performance of the system. And the conventional methods based on characteristics need samples by artificial experiences. But the Characteristics of artificial selection are complicated and difficult to control. Therefore, for automatic extraction of SAR image, essential feature of methods appear to be. Deep learning is a way to automatically extract the image features, but the traditional deep model did not achieve the ideal effect on SAR image because of the complex structure of SAR image. This paper aimed at the above problems is proposed based on the semantic and ridgelet deconvolutional network of SAR image segmentation method, innovation points are as follows:(1) We randomly select a group of basis atoms to initialize the filters of each layer of the deconvolutional network, getting the ridgelet deconvolutional network. Compared with the method of random initializing filters and the method of gaussion initializing filters, ridgelet filters can get better effect on feature extraction, with ridgelet redundant dictionary initialized filter can learn better SAR image characteristics. The SAR image based on semantic space is divided into gathering area, the structure of the homogeneous regions and areas. Of space on the SAR image gathering area of disconnected or homogeneous area training ridgelet deconvolutional network respectively, through the ridgelet deconvolutional network characteristic of the study, said get on behalf of the regional features. However disconnected spatially concentrated areas, homogeneous regions are not necessarily the same kind of ground object, this needs, the initial segmented regions are merged according to the similarity of regional characteristics and through the representative of the regional characteristics of sparse classification, SAR image gathering area and homogeneous region segmentation result.(2) Gathered with regional structural features, homogeneous area have the characteristics of micro texture, area light and shade change is not strong, using the method based on ridgelet deconvolutional network segmentation to get the segmentation result is not ideal, and gray level co-occurrence matrix to processing structure has great advantage: homogeneous area is extracted first samples, to obtain homogeneous area sample entropy of the gray level co-occurrence matrix and secondary statistics, such as correlation calculation sample homogeneous area of gray mean and mean square err or, the gray level co-occurrence matrix of the quadratic statistic grayscale average and variance of a feature vector, to get the feature vector for hierarchical clustering, homogeneous region segmentation results, has obtained the good segmentation effect. Structure features of area generally for boundary, line target and independent, therefore, the structure area using the method based on watershed segmentation, and integrate gathered area, the structure of the homogeneous region and regional segmentation result, eventually SAR image segmentation results are obtained.