Research On Ship Contour Extraction Method In SAR Images Based On Deep Learning

In recent years,synthetic aperture padar(SAR)imaging technology has been developed rapidly,and the efficient use of the rich information contained in SAR images to implement high-precision ship identification has become one of the current hot research problems in the field of remote sensing.Traditional SAR image ship detection methods usually rely on manual assistance or require the construction of complex statistical models,and the detection accuracy and generalization ability in practical applications need further improvement.In this thesis,based on the summary of the research results of SAR image ship recognition based on deep learning at home and abroad,we use the open-source SAR Ship Detection Dataset(SSDD)dataset and SAR-Ship-Dataset dataset as the SAR ship detection image data sources,and focus on the ship detection and contour extraction of SAR images under different scales and scenes.The theoretical research and experimental validation of the methods are carried out.The specific research contents and innovative works are as follows.(1)Joint method with FNLM and Faster R-CNN for ship target detection.Although the deep learning method can obtain a ship detection model with strong generalization ability,there are still problems of missed detection and false detection in some SAR images containing complex scenes.This thesis combines the imaging principle and image characteristics of SAR,and proposes a Fast Non-local Mean(FNLM)filter combined with Faster R-CNN for ship detection in SAR images.First,the FNLM filter is used to filter the SAR ship detection image data;then,the Faster R-CNN is used to perform migration learning on the joint dataset before and after the filtering;finally,the fast detection of the ship target is achieved and the localization information is obtained.The experimental results show that the joint method proposed in this thesis can obtain high ship detection accuracy,which is improved from 0.66 to0.70 compared with using Faster R-CNN.(2)Instance segmentation method for ship contour extraction based on deep learning.In response to the increasing demand for fine-grained detection of ship targets,this thesis uses an instance segmentation method based on deep learning to extract ship contours,using the current representative"two-stage network"Mask R-CNN and"single-stage network"SOLOv2 to segment the ships instances from SAR image to obtain the contours.Among them,the Mask R-CNN is used to extract the ship contour with high target detection accuracy,and the mean value of the bounding box mean average precision(bbox-mm AP)and the mean value of the segment mean average precision(segm-mm AP)reach 0.72 and 0.70 respectively.But the extracted contours had problems such as contour detail loss and contour overlay mask.The extracted ship contour using SOLOv2 has a mask mean average precision(mask-m AP)of only 0.54,and there is a larger scale of missed detection in the small size target group.The experimental results show that although the deep learning-based SAR image ship instance segmentation method can achieve contour extraction,the contour accuracy still needs to be improved.(3)Joint method with Faster R-CNN and optimized Chan-Vese model for ship contour extraction.In order to effectively extract ship contours from SAR images and reasonably evaluate the accuracy of ship contour extraction,this thesis proposes a method to achieve high accuracy ship contour extraction by combining deep learning methods with contour models.First,the Faster R-CNN is trained to achieve fast detection and target area slicing of ships in SAR images;then,the FNLM filter is used to perform background denoised and target enhancement on the slices;finally,the optimized Chan-Vese model is improved to achieve contour extraction,which reduces the impact of noise on the ship contour morphology and reduces the computational effort of the model compared with the original model.In addition,this thesis proposes an index R_N that can evaluate the accuracy of ship contour extraction in two dimensions:bias direction and numerical quantification.The experimental results show that the proposed method achieves an average accuracy of R_N=-0.002 on the SSDD dataset,which is less than the ideal value of R_N=0,indicating that the extracted ship contours are closer to the real contours of the target.To address the problem of how to effectively improve the accuracy of ship contour extraction from SAR images,this thesis,based on the in-depth study of the ship target detection method jointly with FNLM and Faster R-CNN,and the ship contour extraction method based on deep learning instance segmentation algorithm,focuses on the ship contour extraction method jointly with Faster R-CNN and optimized Chan-Vese model,and also proposes the ship contour evaluation index R_N,and the research work in this thesis provides theoretical support and reference for the effective extraction and accuracy evaluation of ship contour from SAR images.