Akademik Veri Yönetim Sistemi
International Conference on Engineering, Social- Sciences, and Humanities (ICESSH-24), Makkah, Suudi Arabistan, 8 - 09 Şubat 2023, ss.1-46, (Özet Bildiri)
Synthetic Aperture Radar (SAR) images are widely used for both military and civilian purposes. These radar images can be captured without being affected by conditions such as fog or darkness, making them ideal for object recognition and detection. In recent years, ship detection from SAR images has become a particularly popular topic. In this study, a machine learning-based method for detecting ships from SAR images has been developed.
First, a new database of 10,000 images was created using the Copernicus OpenAccess Hub, which provides free access to SAR imagery. The images in this database were labeled as either containing ships or not.
A histogram-based method using differential calculations was employed to extract meaningful features from the ship images. The first and second derivatives of the images were calculated using two-dimensional Gaussian functions. From these derivatives, the eigenvalues and eigenvectors of the images were obtained, revealing boundary and edge information in regions containing ships. The Hessian matrix was then computed using these eigenvalue and eigenvector matrices, which plays a critical role in distinguishing between ship areas and the sea surface in the images.
An oriented histogram was calculated on the Hessian matrix, resulting in feature vectors of SAR images with a size of 1x128. This feature vector enhances the classification ability of the model. The nearest neighbor classifier was used to classify the feature vectors, with the Chi-square distance serving as the distance metric in the classifier.
The experimental results showed that ships were detected with a 94% accuracy rate. In future work, the database is planned to be expanded, and ship detection methods using deep learning techniques will be developed.