Research On Infrared Target Detection Based On Domain Adaption

Target detection and recognition methods based on deep networks have achieved great success in recent years and are widely used in various fields.However,the training of such methods requires the support of big data.For military targets,it is difficult to obtain sample data of massive real military targets.For the problem of target detection and recognition under the condition of insufficient sample the sample expansion and transfer learning methods have been widely studied.In this paper,we studied the problem of infrared image target detection and recognitionIn this paper,a sample expansion method based on the time-varying radiation curve of various materials is proposed.Through the combination of material radiation theory calculation and commercial simulation software infrared simulation,the radiation models of various materials in different time periods are established.The radiation contrast relationship of the time period realizes the expansion of the real infrared image in different time periods,and the consistency of the extended image and the real image of the simultaneous segment indicates that the extended data can be used for network training.One method of domain adaptive training for target detection networks is to train different networks with different types of samples.This paper compares the performance of infrared simulation images and visible real images applied to infrared image target detection algorithm training,by using commercial modeling software.Multigen Creator,which models aircraft targets and backgrounds,uses commercial infrared simulation software SEWorkbench to simulate infrared simulation images of aircraft targets in Dual-band,multitime,multi-climate,multi-background,multi-view,and applied to networks.Through the test on the real infrared image,it is verified that the infrared simulation image training detection and recognition network can be better than the visible light image.Aiming at the problem of poor performance of the detection network due to the difference of the characteristics of the source domain sample and the target domain sample,a domain adaptive network with deep confrontation is used.The method of back propagation is used to train the simulated infrared image and the real infrared image.The domain invariant feature is introduced into the Faster R-CNN detection network model,and an end-to-end target detection network structure is established.Simulation experiments show that this method can effectively improve the detection and recognition network based on simulated infrared image training.The true infrared target detection accuracy improves the cross-domain robustness of target detection.