Ship detection from sar images with machine learning and image processing methods

Nowadays, Synthetic aperture radar (SAR) images are widely and extensively used in defense industries, ocean surveillance, environmental monitoring, and emergencies such as disasters. The SAR technology provides high-quality images regardless of the climate and weather conditions, and that feature enables the development of effective object recognition and target tracking applications. In this thesis, two methods have been developed to detect and classify ships in SAR images. Firstly, a thorough SAR ship image database was created. This database includes images from ports, seas, and oceans around the world, thus forming a comprehensive dataset. Four different statistical measurements were initially made from the gray-level co-occurrence matrix to extract relevant feature vectors from the images. Gaussian derivative filters were used to construct the gray-level co-occurrence matrix, and the obtained feature vectors were classified using support vector machines. In the second developed method, the Hessian matrices of the images were calculated. By calculating the eigenvalues of the Hessian matrix, gradient magnitude and gradient orientations were obtained. With this information, adequate histogram labeling was performed, effectively distinguishing between pixel data of ship and non-ship areas. The obtained feature vectors were classified using artificial neural networks. The proposed method was compared with the Local Binary Pattern (LBP) method. Experimental results showed that the proposed methods provided effective results in ship classification.