Ship Target Detection Of Spaceborne SAR Images Based On Joint Full Convolution And Regional Proposal Networks

The detection of ship targets is widely used in civil and military fields such as such as the maintenance of marine rights and interests,and maritime traffic safety.Based on the advantages of Synthetic Aperture Radar(SAR)full-time,all-weather observation and high-resolution imaging,a large amount of high-quality marine SAR data can be obtained,which provides rich data support for sea area detection.Therefore,SAR images are used to complete ship targets detection is of great significance.In order to realize the accurate detection of ship targets in SAR images containing sea and land information,this paper mainly studies the method of sea and land segmentation,and ship target detection based on deep learning technology for GF-3 SAR images.The main work of the paper is as follows:In view of the traditional segmentation methods of SAR images which are usually affected by the specific scenes and the poor robustness,this paper mainly studies the sea and land segmentation algorithms based on the fully convolutional network for GF-3 SAR images.The network uses U-net as the overall architecture,and the residual convolution module is introduced in the encoding and decoding stages to deepen the network structure,and the jump connection method is redesigned for improvement.In addition,large-size original GF-3 SAR images are collected,and dataset expansion and labeling are performed.The traditional segmentation method Otsu,the segmentation method based on U-net and the segmentation method proposed in this paper are used in the sea and land segmentation of the images in test set.The accuracy,mean intersection over union of the segmentation method based on deep learning is significantly better than the traditional method,and the improved method in this paper can better retain the edges of the sea and land in the image.Based on the sacrifice of a small amount of processing time,the proposed method can be exchanged for higher pixel accuracy and intersection ratio to achieve more detailed segmentation of sea and land.Aiming at the problem of inefficient detection of SAR images by Faster R-CNN algorithm,this paper mainly studies the detection algorithm for ship targets in GF-3 SAR images.The feature extraction network,region proposal network,identification network and loss function are introduced.The original high-resolution SAR images are too large,the method of tailoring the region containing the ship targets at different scales is used to facilitate the acquisition of more feature information during the training process,and expanding the dataset of the obtained small-sized images.Finally,the ship targets dataset in VOC2007 format standard.In order to achieve more accurate detection,using K-means clustering ship target size and aspect ratio to improve anchor frame design based on clustering results.By comparing and analyzing the detection effect,the accuracy rate and recall rate have been improved relative to the original algorithm.Aiming at the low recall rate of detection results in large-size SAR images,a method of cropping into small-sized images before detection and then stitching to original size is proposed.By analyzing the detection results,the method of ship targets detection based on joint result of sea and land segmentation is proposed for the problem that some small areas of land areas and targets on the land will be misdetected as ships,thereby making the false alarm rate too high.In addition,the two-dimensional Gaussian function is used to establish the sea surface confidence based on the segmentation results,and the detection results are corrected according to a set threshold to achieve further detection and improve detection precision.The paper implements the sea-land segmentation and ship target detection of GF-3 SAR images,but there are still many problems to be studied.Further optimize the detection algorithm to improve the detection ability of small-sized ship targets.Although the algorithm in this paper can improve the detection accuracy to a certain extent,the detection speed needs to be improved.In addition,this paper only uses the joint algorithm of land and sea segmentation and ship target detection in the test phase,and the model needs to be further refined.