Ship Detection And Recognition In Remote Sensing Images Based On Deep Learning

Nowdays,with the development in remote sensing field worldwide,especially synthetic aperture radar(SAR)and optical sensors,more and more remote sensing data need to be processed.It is hard to process such volume of data manually.This requires related automatic remote sensing image interpretation system.Object detection and recognition is one of the core technology of automatic remote sensing image interpretation system.Based on big data,deep learning has achieved much better performance in object detection and recognition than traditional methods.Since more and more remote sening data is obtained,deep learning based remote sensing image object detection and recognition has become a promising research field.In this paper,we focus on the detection and recognition of ship target in remote sensing images based on deep learning.For problems in traditional SAR ship detection and recognition methods,two improved algorithms based on Faster R-CNN SAR ship detection and two algorithms based on deep learning for SAR ship recognition are proposed.For problems in optical ship detection and recognition methods,an algorithm based on deep learning is proposed.In this paper,based on Faster R-CNN,two deep learning SAR ship target detection algorithms are proposed,progressively.First,because of the miss detection of small ship target and false alarms in detection results,the sequeeze and excitation Faster R-CNN(SE Faster R-CNN)is proposed.Based on Faster R-CNN,the multiscale feature maps are obtained by concated feature maps in different scale to improve detection performance of small ship targets.At the same time,the squeeze and excitation block(SE block)is proposed to generate a weight vector and recalibrating sub-feature maps of each region of interest(Ro I)to suppress redundant feature map channels,which can improve the detection performance.Further more,the rank operation is introduced and the squeeze and excitation rank Faster R-CNN(SER Faster R-CNN)is proposed.Based on the rank operation,the redundant feature map channels are completely suppressed,and the detection performance is improved again.Experimental results based on Sentinel-1 SAR images show that the proposed methods can significantly improve detection performance.For target recognition field,because of relative more labeled data in vehicle target recognition task,a SAR vehicle target recognition method based on convolutional highway unit network(CHU-Net)is proposed.Because of the convolutional highway unit allow for the transfer of information across multiple layers without attenuating any information,high performance deep network can be trained with limited labeled data.The performance of CHU-Net is valided based on Moving and Stationary Target Acquisition and recognition(MSTAR)dataset and used for ship recognition in Sentinel-1 images.At the same time,because of labeled SAR ship data is limited,a SAR ship target recognition method based on GAN pretrained CNN is proposed.The limited labeled SAR ship data is used to train GAN,and the ship images generated by GAN is used to pretrained CNN.Then,the CNN can learn feature of SAR ship target.Finally,the CNN is fine tuned based on limited labeled SAR ship data.Experimental results based on Sentinel-1 SAR images show that the GAN pretrained CNN classifier based on training procedures above can have good classification performance.In high resolution optical remote sensing image ship target detection and recognition task,detecting ship targets in sea area is not challenging.Instead,more effort is needed to detect ship targets in ports because they usually very close to wharf and each other.For current deep learning methods,it is likely to miss detect ships close to each other because of the overlap of bounding box in detection results.In this paper,a rotate object detection method based on PVANet is used for ship target detection.The performance of method is valid based on Google Earth ship detection and recognition dataset and a optical ship detection dataset.