Study On The Recognition Of Ships In SAR Imagery Based On Deep Learning Methods

Automatic ship recognition is an important link of the marine surveillance systems,which has been widely used in fisheries management,marine rescue,traffic control,military operations etc.In particular,Synthetic Aperture Radar(SAR)is often employed in ship recognition due to its ability to penetrate various weather and light conditions.Traditional ship recognition system is divided into three key points,image preprocessing,feature extraction,classification and identification.Feature design and selection is the backbone of traditional recognition methods.On the one hand,pre-specification or pre-selection features manually is a costly and time-consuming effort,which is also prone to errors and omissions.On the other hand,a limitation for man-made features is that they cannot fully reflect high-level information in the SAR imagery.Recently deep learning has achieved state-of-the-art success in many knowledge engineering and shown a great success in the field of optical image classification.Typical deep learning recognition approaches like Convolutional Neural Networks(CNNs)have realized end-to-end classification by learning features automatically.However,the application of CNNs to SAR images are limited for the gap between SAR and optical images,as well as the insufficient annotated SAR data.To address the forementioned issue,the following works in two aspects are made in this dissertation:1)We propose a SAR ship recognition model based on Convolutional Neural Networks.To make our model better fit into the SAR ship recognition task,a series of adjustment and improvement is conducted based on LeNet-5 digital recognition methods.Meanwhile,various methods including training optimization methods,regularization and dropout also apply to our model.Finally,we verify our proposed method through experiment on OpenSARShip database,which contains thousands of real ship slices.Besides,contrast experiments with traditional methods have also been done.2)We propose a SAR ship recognition model based on transfer learning.Considering the lack of annotated samples,our work is to leverage available datasets Moving and Stationary Target Acquisition and Recognition(MSTAR)to extract features that are useful for ship recognition.Specifically,we start with a CNNs model pre-trained for object recognition on MSTAR.Next,we build on the knowledge gained from this image classification task and fine-tune the CNNs on the OpenSARShip.We finally find the best recognition model by fixed different layers of our model.This dissertation validates the rationality and validity of the proposed methods from the points of both theory and practice.CNNs and transfer learning have been demonstrated to be highly informative for the SAR ship recognition,especially showing greater potential for SAR image classification.