Research On Ship Detection Algorithm In SAR Image Based On Optical Domain Adaptation

Ship detection is an important research topic in the field of remote sensing.The performance of optical ship detection is easily affected by weather factors.Compared with optical imaging,Synthetic Aperture Radar(SAR)signal has a stronger penetration ability.And SAR images provide consistent geometric information under all-day,all-weather conditions.This makes it possible to detect ships obscured by clouds and some hidden targets.These advantages make SAR ship detection very attractive.Recently,the state-of-the-art methods exploit convolutional neural networks to train ship detectors,which require a large labeled dataset.However,it is difficult to label the SAR images,which requires expensive labor and well-trained experts.In view of the advantages and disadvantages of the existing two types of ship detection: optical ship detection and SAR ship detection,this paper considers a cross-domain ship detection,that is,the ship detection for SAR images based on optical domain adaptation.It utilizes a large number of labeled optical images to help achieve SAR ship detection.However,the cross-domain ship detection has the following problems: 1)The domain difference between the optical images and the SAR images is large,and the SAR image lacks color information,suffers from speckle noise,and has geometric distortion and shadows.2)The supervision information of optical images and SAR images interfere with each other,and the labels of a large number of optical images tends to bias the detector to the optical domain and ignore the SAR domain.Therefore,the main research contents of this paper are as follows:(1)To narrow the domain shift between optical images and SAR images,this paper proposes a cross-domain ship detection algorithm with a multi-level alignment network.It aligns the distribution differences between the optical domain and the SAR domain from the image-level,convolution-level,and instance-level,respectively.First,image-level alignment exploits generative adversarial networks to generate SAR images from the optical images.Then,the generated SAR images and the real SAR images are used to train the detector.To further minimize domain distribution shift,the detector integrates convolution-level alignment and instance-level alignment.Convolution-level alignment trains the domain classifier on each activation of the convolutional features,which minimizes the domain distance to learn domain-invariant features.Instance-level alignment reduces domain distribution shift on the features extracted from the region proposals.The entire multi-level alignment network is trained end-to-end and its effectiveness is proved on multiple cross-domain ship detection datasets.It achieves 57.37% m AP on the HRSID dataset and 61.26% m AP on the SSDD dataset by utilizing labeled optical images.(2)To avoid the mutual interference of optical image and SAR image supervision information,this paper proposes a cross-domain ship detection algorithm based on a semi-supervised co-training framework.this paper proposes to decompose different supervision sources to form two sub-tasks: cross-domain and semi-supervised.The Dual Teacher framework is proposed,which contains two co-trained teacher-student models.One learns cross-domain consistent knowledge using the cross-domain dataset consisting of labeled optical images and unlabeled SAR images.Another learns SAR domain unique knowledge using the semi-supervised dataset consisting of labeled SAR images and unlabeled SAR images.Non-maximum suppression is used to fuse the predictions of the two teacher networks and generate high-quality pseudo-labels to guide the better training of the student networks.These two kinds of complementary knowledge make the detector achieve better performance.The effectiveness of the Dual Teacher framework has been fully experimentally demonstrated.It achieves 68.5% m AP on the HRSID dataset and 82.5% m AP on the SSDD dataset by utilizing labeled optical images and three SAR sample labels.