Fast ship detection combining visual saliency and a cascade CNN in SAR images

Abstract

In order to realise the fast detection of ships in synthetic aperture radar (SAR) images, a detection method combining visual saliency and a cascade convolutional neural network (CNN) is proposed. First, based on visual saliency, a multiscale spectral residual model is designed for realising the fast detection of ship candidate regions. Then, a cascaded CNN is designed, which consists of two convolution networks, namely, the front-end shallow CNN, which is used to quickly exclude obvious non-ship candidates and classify the ship candidates according to the ship orientation, and the back-end deep CNN, which is used to detect high-probability candidate regions with rotatable boundary boxes. The whole structure can realise the fast detection and precise positioning of ships with an arbitrary orientation. Finally, the authors conduct detailed experiments on the SAR ship image dataset. The experimental results show that the proposed method can effectively improve the detection accuracy of ships, ensuring the detection efficiency in SAR images.