Research On Ship Detection In SAR Images Based On Deep Learning

Object detection is widely used in human-computer interaction,security monitoring,unmanned driving,aerospace,marine and many other fields.With the rapid development of the maritime transportation industry in recent years,higher requirements have been put forward for the detection of marine ship objects.Ship detection combines the ship classification task and ship localization tasks.Ship classification is responsible for distinguishing the foreground and the background,and ship localization is responsible for marking the position of every ship with the bounding box.And the accuracy and the speed of detection are the key factors in the task of maritime ship detection.The main sources of marine ship images are Synthetic Aperture Radar(SAR)images and optical remote sensing images Different from natural images,SAR images are independent of sunlight and weather conditions,so SAR images are widely used in maritime ship detection tasks.Traditional ship detection methods in SAR images often lead to poor robustness of the model.In order to obtain more accurate results of ship detection,based on deep learning with strong representation learning ability,this thesis focuses on ship detection in SAR images and improves it,so as to optimize the ship detection effect in SAR images,which has certain significance for the research of ship detection algorithm with higher detection performance.The main research contents of this thesis are as follows:(1)Object detection algorithm based on deep learning mainly contains one-stage detection algorithm and two-stage detection algorithm.Firstly,this thesis is based on the representative algorithm of the two-stage object detection algorithm Faster R-CNN,the multi-stage object detection algorithm Cascade R-CNN,and the representative algorithms of the one-stage detection algorithm RetinaNet and Libra RetinaNet have carried out the research of SAR ship detection,in order to determine the better detection algorithm type in SAR ship image.(2)The object detection algorithm based on deep learning is driven by data to a certain extent,that is,the richer the data set,the better the result.Due to the insufficient amount of original SAR data set and the limited complexity of the image background,this thesis has carried out research on effective data enhancement strategy around SAR ship images and proposed a data enhancement strategy for SAR ship images based on style transfer to expand the data set.In order to obtain a richer SAR images ship data set,so as to improve the detection performance of the detection model on the SAR image ship data set.(3)A great challenge currently facing ship object detection in SAR images is the problem of small object detection.In the process of feature extraction,low-level features mine detailed information,which contains more information about edges and contours.This information is of great significance for small object detection.In addition,the one-stage object detection algorithm can guarantee a relatively high detection accuracy while ensuring a relatively fast detection speed,so this thesis proposes the improved Libra RetinaNet to reduce the loss of low-level features in the feature transfer process.The improved Libra RetinaNet mainly improves the feature fusion module in original Libra RetinaNet to better save and transfer low-level features.The improved Libra RetinaNet achieves 97.38%mean Average Precision(mAP)on the SAR test set.