Research On Ship Detection Algorithm In SAR Images Based On Deep Learning

In recent years,with the continuous development of deep learning in the field of target detection,technology of SAR images target detection has been widely used in ocean monitoring,fishery management,scientific research and other fields.At the same time,we notice that most of the existing detection models are based on optical images,which can not achieve the expected results when applied directly to ship detection of SAR images,and SAR images is more difficult to obtain than optical images,so there are usually insufficient data in the process of model training,leading to the problem of model over fitting.Therefore,this paper mainly studies the SAR images generation algorithm based on depth learning and the ship target detection models suitable for SAR images.The purpose of this paper is to solve the problem of insufficient data in the process of SAR images target detection by using images generation algorithm based on depth learning,and to realize the models suitable for ship detection of SAR images by combining features fusion,double attention mechanism and other technical means.The main work of this paper includes:(1)The paper proposes a deep convolutional generative adversarial networks based on Wasserstein distance.To expand the data set of SAR ship detection and enhance the diversity of data,this paper presents a generative adversarial networks model WI-DCGANs,which combines DCGANs and WGANs.Aiming at the problems of unstable model training,insufficient diversity of generated samples and low quality of traditional generative adversarial network,the generator and discriminator based on depth neural network are proposed.In the generator model,the idea of residual is introduced to avoid the problem of gradient disappearance on the basis of increasing the depth of network and improving the ability of model feature extraction.At the same time,Wasserstein distance is introduced into the loss function of the model,and the gradient penalty strategy is used to limit the range of data generation to speed up the convergence of the model.Experiments show that the generated adversary network can effectively generate SAR images and enrich data set.(2)The paper proposes an improved SSD target detection algorithm TF-SSD based on multi-scale feature fusion.In view of the fact that the size of ship target in SAR images is generally small,this paper introduces the method of deconvolution network and element sumfeature fusion based on SSD network model,and uses the shallow feature map and the fusion information among feature layers to make up for the disadvantage that SSD network is not robust enough to small ship target in SAR images.Compared with SSD model,the mean average precision of the proposed model is improved by 2.6%;compared with Faster RCNN and YOLOv3,the comprehensive evaluation index F1-score of the proposed model is improved by 5.5% and 7.9% respectively,which shows that the model has good robustness for SAR ship detection;at the same time,the technology of data enhancement can effectively improve the detection performance of the model.(3)The paper presents a ship detection algorithm DASSN based on dual attention mechanism for complex background SAR images.Aiming at the problem that ship detection in complex background SAR images is easily affected by ground objects,which leads to low detection rate of model,the dual attention mechanism of channel and space is introduced into the target detection network;the expansion convolution and concat feature fusion technology are applied to the target detection network to improve the robustness of model to small targets;in order to improve the detection speed of model,the lightweight Mobilenet is used;and a new two class loss function is used to make the model training set different weights for difficult and easy samples.The experimental results show that the mean average precision and F1-score of the proposed algorithm are 88.9% and 91.2% respectively,and the detection speed is 42.1 fps,which proves that the proposed model can not only effectively improve the detection accuracy of SAR images with complex background,but also improve the detection speed to a certain extent.In conclusion,the proposed WI-DCGANs model can effectively generate SAR images,enhance data sets,and provide data basis for model training;TF-SSD network can better apply to SAR images ship detection,especially for small-scale ship detection;DASSN can effectively improve the ship detection effect of the model on SAR images with complex background.