Ship Target Detection And Recognition In Radar Images Based On Small Samples

With the continuous development of modern technology,ship target detection and recognition algorithms based on radar images are also emerging in endlessly.Although the classic deep learning target detection and recognition algorithm has a good performance in accuracy,its detection rate still does not meet the requirements of many military applications.Moreover,due to the particularity of radar images,it is difficult to obtain a large number of radar image samples in practical applications.Therefore,the detection and recognition of ship targets in radar images based on small samples is a challenging and meaningful research topic.Aiming at this issue,this paper adopts the method of generating countermeasure network combined with ship target’s three-dimensional electromagnetic scattering information to simulate the ship target’s radar image and expand the sample generation based on the depth generation network,so that the expanded ship radar image is not only better than the traditional method The expanded image is closer to the real image and greatly increases the diversity of the samples.Finally,an end-to-end convolutional neural network based on regression is used to extract features of different depths of ship targets to detect and recognize ships,and to improve the detection accuracy and generalization performance of ship targets in the case of small samples.The main research contents of this paper are:1.Research on the radar image data expansion method based on the generative countermeasure network.This article introduces the generative countermeasure network into the field of radar image generation,which can generate artificial radar image samples that are very similar to real samples.After learning,the generative countermeasure network can generate some "forged" ship images with different azimuths and angles to expand the radar image database,and achieve the purpose of alleviating the lack of radar image samples that affects the accuracy of ship target detection.2.Aiming at the problem that traditional data augmentation methods cannot truly expand the image angle and orientation of the image,this paper studies a small sample expansion method based on electromagnetic simulation of three-dimensional models.First,based on the ship’s three-dimensional model and backscatter characteristics,the radar image simulation of various ships under different angles and poses is carried out to achieve the purpose of expanding the radar image data.Then combined with the generation of confrontation network,the image is further expanded more realistically,and finally the endto-end deep detection network is input to complete the detection and recognition of different targets.3.Aiming at the problem that the classic two-stage target detection and recognition algorithm is inefficient and difficult to meet the actual application needs,this paper studies the singlestage target detection algorithm,and models the target detection and recognition task as a regression problem.The end-to-end single-stage convolutional neural network is used to achieve rapid detection and recognition.At the same time,by improving methods such as data enhancement,introducing feature pyramid network,path aggregation network and other modules,the model can extract and integrate the features of different depths of the target,so as to realize the radar image Accurate detection and identification of ship targets.