False Alarm Source Detection And Recognition Based On Migration Learning For Earth Observation

Earth observation data is an important data source in the field of remote sensing detection,and infrared band data is free from sunlight interference,so it is of great significance to be used for missile trajectory detection,early warning and interception in the field of military and national defense.However,in the infrared imaging band of the earth observation system,there are a large number of false alarm sources with the same radiation characteristics as the detection and tracking targets in the military and national defense field,which will cause interference to the missile tracking and early warning system and improve the false alarm rate of detection.Common false alarms include forest fires,high cirrus clouds,snow-capped mountains,frozen lakes and islands.In order to reduce the false alarm rate and improve the accuracy of subsequent target detection,it is necessary to detect the false alarm source to remove the interference in the infrared satellite image.With the wide application of deep learning in the field of computer vision,the depth model has the characteristics of generalization of detection tasks and more robustness compared with traditional image processing target detection,which is convenient for efficient detection of false alarm sources.Due to the sensitivity of earth observation satellite data in national defense and the characteristics of infrared remote sensing imaging,it is difficult to obtain data sources,and the quality of existing data is relatively high.In this paper,the deep learning target detection method is mainly used to study infrared image preprocessing,false alarm source feature extraction network and false alarm source detection and recognition technology,and carry out theoretical investigation,method improvement and comparative experiments to solve the above problems.The main contents of this paper are as follows:1、The basic theoretical knowledge of deep learning target detection based on transfer learning is studied,including the foundation,classification and application of transfer learning,the basic structure of convolutional neural network,commonly used feature extraction network and target detection framework;2、In view of the problem that the local contrast is too low caused by the mist in the infrared satellite image of earth observation and the imaging characteristics,the pretreatment methods of infrared satellite image of earth observation are studied,including the fog removal algorithm based on visual salience and dark channel,and the histogram equalization method based on the contrast limitation;3、In view of the scarcity of infrared false alarm source data,the source domain data that is most suitable for transfer learning is studied and analyzed,and the feature extraction network pre-training method based on parameter migration is completed,and a comparative experiment is designed to verify the improvement of transfer learning on the detection performance of false alarm source;4、In view of the point cloud clusters of volume and the characteristics of the source of false alarm,combined with the false alarm source multi-scale,uneven distribution of features,studies the basic network,attention mechanism,the feature fusion method,the classification of return loss function improvement methods,and puts forward the source of false alarm detection based on migration study to identify network,improved under the condition of limited data source of false alarm detection performance.