Ship Detection from Despeckled Satellite Images using Deep Learning

Abstract

Synthetic Aperture Radar is a technique of forming fine resolution images from resolution restricted radar systems. SAR images are captured by a moving system such as an airplane or a satellite using the motion of a radar antenna. The most common application of SAR images is object detection. In this paper, a ship detection model has been proposed from satellite images using deep learning. In general, a SAR system generates speckled images, which can reduce the performance of a detection model. To reduce deleterious effects, various speckle filtering methods have been used by several researchers in the past. De-speckling is a technique which reduces noise (speckle) from the images. In this paper, two filters are used for despeckling the SAR images and the images are used for evaluation of the proposed Convolutional neural networks. The proposed CNN based deep learning model has been used to classify the object present in the images as ship or not-ship. A comparison has also been made between speckled dataset and despeckled dataset as well as for different epochs in the model to find the best results.