Research On Ship Target Detection In Optical Remote Sensing Image Based On AR-RPN

With the development of globalization,the demand for marine transportation and coastal defense construction is growing rapidly.Ship target detection in remote sensing images is playing an increasingly important role in port monitoring,ship management,maritime rescue,and coastal defense construction.The rapid growth of maritime traffic also makes ship target detection face greater challenges.The development of deep learning research has made convolutional neural networks more and more applied to the field of remote sensing images.At present,in the process of ship detection research on optical remote sensing images,extracting ship targets from complex backgrounds is the bottleneck of ship detection research.The large differences in the shape distribution,arrangement distribution,angle distribution and scale distribution of ship targets in remote sensing images with overhead angles have led to the low detection accuracy of conventional target detection algorithms on optical remote sensing image ship target data sets.Therefore,on the basis of the two-stage target detection network,this paper constructs two ship target detection networks with adaptive rotating region proposal network as the core by improving the feature extraction network,feature fusion pyramid network and region proposal network and other components.And through multiple sets of experiments,the effectiveness of this article to improve the network is verified.The main research work of this paper is as follows:Work 1: Proposed Adaptive Rotating Region Proposal Convolutional Neural Network(AR-CNN)for ship target detection in remote sensing images.AR-CNN uses an improved multi-scale feature pyramid network to perform feature fusion to generate feature maps,and combines with an improved adaptive rotating region proposal network to generate candidate frames for ship detection.A combination of anchor frames that are more suitable for the length and width ratio of the ship is designed,so that the network can effectively detect ship targets arbitrarily distributed in the remote sensing image.Finally,two kinds of loss,anchor box position loss and anchor box shape loss,are added to constrain to speed up the network convergence.In the experimental part,after ablation experiments on different components of AR-CNN,the effectiveness of the improved network is verified.Comparing experiments with other methods on the DOTA and HRSC2016 data sets,the detection accuracy of AR-CNN reached 87.2% and 88.3%,respectively,which further proved the effectiveness and generalization ability of AR-CNN.Multiple experiments have shown that the improved components of AR-CNN can increase the detection rate of ship targets compared to the basic target detection network.Work 2: Constructed a Multi-Path Feature Extraction Convolutional Neural Network(MP-CNN)for multi-scale ship detection.In response to the problem of ship’s intra-class scale diversity,AR-CNN enhanced the detection performance of small ships by improving the feature pyramid for feature fusion,but there is still a large room for improvement for the multi-scale ship target detection problem.MP-CNN adjusts the receptive field through expansion convolution,and sets different expansion rates corresponding to different scale targets.Through methods such as target scale matching and feature weight parameter sharing,the detection rate of targets of various scales can be improved without increasing the amount of calculation.Aiming at the irregularity of the overlapping area in the oblique frame detection process,the method of rotating non-maximum suppression is improved.Through the calculation of the center point distance constraint,more matching candidate frames are selected during the non-maximum suppression process,which is more conducive to network learning.Through ablation experiments on the DOTA data set,multi-branch and multi-scale comparison experiments,and comparison experiments with other methods on the DOTA and HRSC2016 data sets,the detection accuracy reached 88.1% and 88.9%,respectively.MP-CNN is effective Performance and robustness have been verified.