An Unsupervised Ship Classifier for High-Resolution SAR Images

This paper presents an unsupervised ship classifier for high-resolution synthetic aperture radar (SAR) images. Firstly, the algorithm to extract the geometric features of ship targets from high-resolution SAR images is described. Since conventional edge detectors can hardly get the correct ship profile affected by speckle noise and strong scatter point in SAR images. A new method combining the ratio-of-averages edge detector and image morphological processing is presented to extract ship profile. With the ship profile, some other geometric features are also extracted including a new ship feature representation based on the space distribution of strong scatter points. Secondly, ship texture feature extraction and dimension reduction are discussed in terms of gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP). Finally, the ship classification model is derived with the expectation maximization (EM) algorithm of unsupervised learning according to the ship features. Ship images from TerraSAR-X have been used as training and testing data, and experiment results are given, which proves the good performance of the classifier.