Fast Detection Of Ship Target For Large Scale Remote Sensing Image Based On Deep Learning

With the rapid development of remote sensing image technology,object detection has become one of the key technologies in the field of remote sensing.Due to the powerful feature extraction capabilities and end-to-end learning style,deep learning has become one of the most effective methods for target detection from large-scale high-resolution remote sensing image.To the end of fast ship detection for large-scale remote sensing image,this paper designs a two-stage detection strategy(reduces redundant calculations brought by the sliding window method).Furthermore,to achieve anchor-free target detection,key point estimations are added to the deep learning network,which helps to obtain faster detection.The main innovations in this article are:(1)A cascade fully convolutional neural network is designed to achieve regionally focused ship detection,which consists of a pre-screening network(P-FCN)and a target accurate detection network(D-FCN).P-FCN is a lightweight image classification network for focusing on target areas.D-FCN realizes the fine positioning of ship targets with arbitrary orientation by adding target mask and ship orientation estimation layer in the traditional U-Net structure.Experiments on Terrasar-X images and optical remote sensing images obtained from Google map show that the proposed network achieves 3 times faster detection than sliding window ship detection method,keeping almost the same level in accuracy.(2)In order to further improve the detection speed,an anchor-free target detection network(RA-CenterNet)based on key point detection is designed.This method predict the target center point and the target size(including length,width,and angle)at the center point via feature learning,location learning,and network training based on course learning.Experiments on SAR images and optical remote sensing images show that the RA-CenterNet network achieves 1.58 times faster detection when comparing with D-FCN,while keeping almost the same level in accuracy.(3)In order to further verify the effectiveness of cascade fully convolutional networks in practical applications,a cascade network object detection system based on a combination of software and hardware is designed.By encapsulating the P-FCN network and RA-CenterNet network in the system,the ship target can be quickly detected in large-scale remote sensing images.In the experiments,SAR images and optical remote sensing images were used to verify the performance of the prototype system.The experimental results show that the prototype system can maintain a certain detection accuracy and detection speed in target detection,which plays a significant role in real applications.