Deep Learning Based Ship Detection For Synthetic Aperture Radar Images

Object detection technique which deals with recognizing and locating certain objects in images,is one of the key tasks of image understanding and it has always been a hot topic for both domestic and international researchers.With the development of satellite technique,especially the launch of Gaofen-3 satellite,a large number of synthetic aperture radar(SAR)image can be used for dynamic sea monitoring which contributes to the development of ship detection for SAR images.In contrast to the object detection of nature images,the ship detection for SAR images has its own characteristics such as low image resolution and meagre image feature.When facing SAR images from different sources,the existing SAR ship detection algorithms present a non-robust detection performance for ships with different sizes and lead to high false alarm rate on land areas.On the other hand,deep learning technique which adopts neural networks to intelligently exploit the features of images in a data-driven way to successfully complete the task,has receives rapid development in the field of computer vision.Motivated by this,this thesis adopts the deep learning technique to study the ship detection for SAR images.The main contributions are as follows.Firstly,to address the high false alarm rate in the land areas,we propose a novel sea-land segmentation based ship detection algorithm for SAR images.The proposed algorithm adopt convolutional neural networks(CNNs)to develop two modules:sea-land segmentation and ship detection.The sea-land segmentation module is used to extract the land areas in the images pixel-wisely;The ship detection module is designed to detect all the ship objects in the images.The proposed algorithm finally merges the results of these two modules to remove the false alarms to obtain the final results.Secondly,experimental results on the public SAR ship detection dataset SSDD indicate that the proposed LDSSD algorithm in this paper can effectively reduce the false alarm rate in the land areas and can be adapted to different SAR images of different resolutions and sources.In addition,the proposed algorithm presents the robust detection performance for multi-size ship objects.Specially,to train the proposed sea-land segmentation ship detection algorithm,we build the first sea-land segmentation dataset on SAR images(SLSS).The SLSS dataset consists of 1270 images including SAR images of different resolutions,sizes,sea environments and sensor types and their corresponding groundtruths,which presents the variation and variety of ship images and provides efficient support of this research field.The above work makes progress in improving detection accuracy of the ship detection for SAR images,enriches the fundamental of ship detection and enhances the availability of ship detection for SAR images.