Ship Detection And Classification Of SAR Images Based On Deep Learning

Synthetic aperture radar(SAR)has become the main means of ocean monitoring.The ship detection and classification of SAR images are of great significance to the management of marine fishery,the management of marine traffic,the fight against illegal sea activities and maritime search and rescue.Deep learning can adaptively extract multi-level and multi-scale deep features,and the detection and classification process can be realized end-to-end.For the huge application needs of maritime trade,traffic management and maritime safety,in order to make full use of the advantages of deep learning and tap its potential in ship detection of multi-scale SAR images and fine-grained ship classification of high-resolution SAR images,this paper carries out ship detection and classification of SAR images based on deep learning.The main research contents and innovations include:(1)Aiming at the problem of ship detection in multi-scale SAR images,considering the detection speed and accuracy,a ship detection framework of multiscale SAR images based on improved YOLO V3 is proposed.To deal with the multiscale characteristics of remote sensing images,YOLO V3 target detection network is built,and multi-scale feature prediction units are designed to achieve multi-scale prediction of ships.The SAR ship dataset is expaned and improved,and the anchors of feature maps are optimized based on large-scale lables clustering to increase the positioning accuracy of target prediction.A backbone network with less calculation is used to extract basic features to reduce the detection time of the whole SAR image.By analyzing the influence of super parameters on the training results,the model with the highest average precision on the test set is obtained.The experimental results on eight SAR images with different resolutions show that: 1)in the open sea area,the detection rate of this method can reach up to 98.2%,in the SAR images with complex sea conditions,the detection rate of this method can reach up to 94.6%,2)based on the optimized YOLO V3 network can accurately detect multi-scale ships.3)compared with other models,the proposed method improves the detection rate of small-scale ships and reduces the false alarms by at least 4.5%.(2)Aiming at the problem of ship detection in heavy sea state SAR images,a ship detection method based on multi-level feature pyramid network is proposed.In the heavy sea condition,the background clutter is increasing,and the contrast between the ship targets and the sea background is becoming low,which makes the ship detection of SAR images become a difficulty in the heavy sea condition.In this paper,we analyze the imaging characteristics of SAR images of sea surface and ship under heavy sea condition.A multi-level feature pyramid network is designed,and multilevel and multi-scale information are added into the prediction model to effectively model the ship target in heavy sea condition SAR images.Based on the trained M2 det model,the whole image of the test image is detected and compared with the detection results of another model.The experimental results show that the algorithm can maintain 90% detection rate and the false alarm rate is less than 10% in heavy sea state SAR images.Compared with other algorithms,it increases the detection rate of small ship targets in heavy sea conditions,and effectively reduces the false alarm rate on land through multi-level and multi-scale feature fusion.(3)Aiming at the problem of fine-grained ship classification algorithm in highresolution SAR images,a fine-grained ship classification method based on deep residual network is proposed,including the analysis of the impact of incident angle factor on the classification results.A multi-source commercial ship dataset of highresolution SAR images is built,and the number of the three categories of commercial ships is balanced and a wide range of incidence angle coverage is maintained.The output of the activation function is still maintained by building residual modules and applying batch normarlization.A deep residual network is build to extract the finegrained characteristics of the ships;and three different fine-tuning strategies are designed to select the best training scheme suitable for fine-grained ship classification of high-resolution SAR images.Data augmentation is used to solve the over fitting problem of small sample training,and the influence of the training set covering different incident angles on the final classification results is analyzed.The experimental results show that: 1)the overall accuracy of the proposed deep residual network is better than 95% in the final ship dataset of SAR images,which proves the effectiveness of the proposed method.2)The incident angle is an important factor affecting the ship feature extraction.The coverage of incident angle of training set samples affects the generalization performance of training model.