The Research On The Applications Of Generative Adversarial Network In Radar Image Semantic Segmentation And Target Detection

Radar(Radio Detection And Ranging)is a device that uses radio to detect targets and acquire their location information.It has very important application value in military and civilian fields.With the continuous maturity of radar imaging technology,rapid and effective automatic interpretation of radar images has very important research value.Radar image semantic segmentation and radar image target detection are important branches of radar image interpretation technology.The quality of radar image semantic segmentation and target detection is the guarantee for subsequent high-quality and high-efficiency radar image processing and interpretation.In recent years,deep learning has been widely used in radar image interpretation technology due to its excellent feature extraction capabilities.As a new star in the field of deep learning,Generative Adversarial Network(GAN)has a unique confrontation game idea that can achieve excellent performance in most computer vision tasks.Therefore,it is of significant research value to introduce GAN into the field of radar image interpretation.This thesis will conduct in-depth research on the application of GAN in radar image segmentation and target detection tasks.The main works of the thesis are as follows:1.Research the semantic segmentation method of radar image based on GAN.In order to ensure the performance of radar image semantic segmentation under the condition of imprecise semantic annotation,a radar image semantic segmentation method based on Pixel to Pixel Image Translation with GAN(Pix2pix GAN)is studied.The research method uses radar images as conditional information to guide the direction of data generation,and On the basis of the generator's encoding-decoding architecture,skip connections are introduced to improve the performance of the model in segmentation details.The discriminator uses Markov discriminator to help the model construct the local information of the segmented image,and further improves the performance of the details.In addition,the L1 loss function is introduced to construct the global information of the segmented image,so that the overall performance of the predicted segmented image is better.The experimental results based on the simulated ISAR image data set show that the research method can achieve better segmentation results,and in the case of insufficient semantic annotation,the model can still achieve more accurate segmentation,compared with traditional deep semantic segmentation methods the model is more robust.2.Research the SAR image target detection method based on adversarial domain adaptation.Compared with optical images,SAR images are more difficult to acquire.Moreover,SAR images are sensitive to radar parameters,imaging modes,grazing angles,etc.,which makes SAR image target detection face a more serious problem of small training data size.In the task of small training data size,the results of target detection on radar images using deep network are usually not ideal.In order to improve the accuracy of SAR image target detection with small training data size,a SAR image target detection method based on adversarial domain adaptation is proposed.The proposed method uses the target detection network Faster R-CNN as the basic framework to construct a dual-stream network with optical images as the source domain and SAR image as the target domain.A domain adaptation module based on GAN constraints is introduced between the two networks,then learn transferable features and realize instance-level domain adaptation by the domain adaption module.In this way,the auxiliary learning of the optical image on the SAR image is realized,and the over-fitting problem caused by the small training data size of the SAR image is alleviated,thereby improving the target detection performance of the SAR image.The experimental results show that compared with the existing methods,the proposed method has higher detection accuracy in the task of SAR image target detection with small training data size.