Research On Multi-Modal Ocean Target Data Generation Method Supporting Small Sample Learning

With the prosperity and development of the maritime industry,various hidden maritime security problems have gradually emerged.As a land-sea complex country,my country’s maritime security is particularly important.For the sake of maintain sea area safety,it is necessary to speed up the algorithm of target detection and recognition,multiple modal data fusion,anomaly warning and threat analysis.However,it requires a large amount of real data samples of sea area sensors to train the above algorithms,and the quality requirements of the data set are high.At present,although there is a lot of information collected by marine monitoring,these data contain sensitive information,and the number of public data sets is limited and the speed is slow.Most of the collection work focuses on a single means,and the data is poorly correlated across time and space.In real application scenarios,the parade targets and the monitoring equipment are always in normal working state,and it is difficult to collect abnormal data.In view of the characteristics of small samples with small amount of marine data,unbalanced distribution and insufficient scenario coverage ability,the main work of this paper includes four parts:(1)An image generation method of water photoelectricity based on least square generative adversarial networks is proposed.By upgrading the discriminant network to multi-scale discriminant network and carrying out high resolution reconstruction,the photoelectric image with higher quality and more realistic details can be generated.(2)An underwater vehicle track generation method based on transfer learning is proposed.Two groups of variational auto-encoders are used to encode and decode the data of the ship’s trajectory and the underwater vehicle’s trajectory respectively,and a gate recurrent unit is introduced in the training process.In the form of cross structure,the simulation track with real sample is generated to achieve the purpose of data increment.A data simulation method for abnormal situation sea area is proposed.Define and classify different abnormal situations,and adopt the method of double probability joint distribution to realize the generation of abnormal data.Based on the track of the submarine,the point track of the working state of sonar and radar sensor is simulated.(3)Design and implement sea area monitoring and management system.With the above data generation algorithm as the core,it realizes the multi-modal sea area target data generation module and the sea area situation monitoring module,and the attribute definition and abnormal situation simulation for the parade targets and monitoring equipment are completed,providing a visual platform for generated data display and multi-modal data fusion.