SAR Image Classification Based On Adaptive Autoencoder And Superpixels

Synthetic Aperture Radar(SAR) can be used to create images of objectives on the surface of the earth and even can penetrate the earth's surface. Due to its characteristic of high spatial resolution, high-resolution SAR images are playing an increasing important role in military affairs, agriculture and medical science. In recent years, SAR image categorization has been widely used in civil and military fields. The current challenge faced with high-resolution SAR image is to extract features of informative richness, strong identification capability and noise suppression. Although previous researchers have made much contribution to it, but the most reliable feature is yet to be found.Therefore, the paper proposes a categorization approach of SAR images based on Gamma distribution and an adaptive multi-scale sparse auto-encoder, as well as a heterogeneous superpixels algorithm which can be effectively used for categorization, segmentation, preprocessing and post processing of SAR images. Specific improvements are listed as follows:Frist, Superpixels algorithm can capture redundant information and reduce the complexity of subsequent tasks, and is a preprocessing method widely used in SAR image categorization and segmentation. But due to the excessive multiplicative speckle noise and frequently appeared weak edge zones of SAR images, traditional superpixels algorithm cannot effectively extract homogeneous pixel blocks and will cause severe interference to subsequent categorization and segmentation algorithm. This paper presents a Gamma-distribution-based heterogeneous superpixels algorithm for SAR image segmentation which is applicable to preprocessing of SAR image categorization and segmentation.Second, Since SAR images bears excessive multiplicative speckle noise as well as homogeneous, non-homogeneous and extremely non-homogeneous zones, a traditional sparse auto-encoder which extracts the pixels characteristics training of single image with fixed scale now is not applicable to effectively present characteristics of high-solution SAR image. The SAR image categorization method proposed in this paper will estimate each pixel and the heterogeneity of its neighborhood based on Gamma distribution when extracting label-free image patches. Then it will adaptively extract various image patches based on their different heterogeneity and train the characteristics before effectively presenting the characteristics of each pixel.Last, Since characteristic training of pixel conducted by a sparse auto-encoder lacks of consideration of spatial neighborhood information, the paper attempts to propose an auto-encoded categorization method of high-solution SAR image which combines deep characteristics and improved spatial information. It can be used in the superpixels algorithm to limit the categorization results from a spatial perspective. The experiment result shows that the categorization method based on the heterogeneous superpixels segmentation and adaptive, multi-scale sparse auto-encoder can categorize high-solution SAR images in a highly accurate manner.