Research On Transfer Learning Method For Ship Target Recognition Based On Optical Remote Sensing Image

Ships are the focus of research on maritime target recognition.The detection and identification of ships has always been a hot research topic in the field of remote sensing images.With the increase in the number of satellites,the world's remote sensing data collection trends are moving toward multi-platform,multi-angle and multi-sensor directions,and the number of satellite remote sensing images obtained is also increasing.Although the target data acquired continues to increase,there is very little data labeled.Since there are some differences between images from various satellites,it is important to study how to use a small number of labeled samples to identify typical targets with different resolutions and different imaging angles.Transfer learning can use the knowledge learned from existing sample data to help sample data in differently distributed unknown areas.Therefore,this paper aims at optical remote sensing images,and uses the transfer learning method to study the methods of three typical ship target recognition.Firstly,the multi-dimensional feature extraction of ship targets is studied.For the optical ship image,the geometric characteristics of the target(Hu moment and affine invariant moment)are studied.For different resolution images,the histograms of oriented gradient features are studied.The differential attribute profile features are studied for different angle ship targets.The two features and geometric features are used to construct multi-dimensional feature vectors for ship recognition,and the maximum mean discrepancy is obtained.The maximum mean discrepancy distance and transfer component analysis methods are studied to analyze the feasibility of ship recognition using transfer learning.Secondly,the target recognition method based on feature based transfer learning is studied for ship image recognition problems with different resolutions.Firstly,a transfer learning method based on data distribution adaptive,namely spatial alignment-joint probability adaptation algorithm,is proposed,which combines spatial adaptation and probability distribution to achieve better recognition results.In addition,the mapping statistical alignment algorithm is also studied.It adopts a unified framework to increase the distance between the source and target domains when the distribution is aligned,and reduce the intra-class distance to make the ship recognition more accurate.Through experimental verification,it can be found that for different resolution ship targets,the transfer learning algorithm can effectively reduce the distribution difference between the source domain data and the target domain data,thus achieving better recognition effect than the traditional machine learning method.Finally,the target recognition transfer learning method based on instance based transfer learning is studied for the identification of ship images at different angles.Firstly,when there are part labels in the target domain,a supervised transfer learning method based on Mahalanobis distance is studied,which mainly increases the weight of similar samples in the training sample and the target domain to achieve to achieve better recognition results.Aiming at the fact that there is no label in the target domain and the difference between the source domain and the target domain is large,the unsupervised transfer learning method called transfer joint adaptation method is studied.It mainly combines instance transfer and probability adaptation,and verifies the effectiveness of the algorithm through experiments.