| New types of targets with low observability such as small unmanned aerial vehicles and stealth aircraft present major challenges for modern radar systems.Due to their extremely low radar cross-sections and their tendency to be obscured by complex land and sea clutters,traditional target detection methods often result in severe false alarms and missed detections.Thereby,this thesis explores intelligent radar target detection methods under nonhomogeneous,time-varying clutter conditions,proposes new methods based on complex-valued neural networks for radar clutter suppression and target detection,and develops a radar self-evolving target detection architecture suitable for few-shot conditions,significantly enhancing radar target detection performance in complex time-varying clutter environments.The main contributions and innovations are as follows:(1)A nonhomogeneous clutter suppression method based on conditional generative adversarial network is proposed.By using adversarial training,a generator capable of extracting target signals from complex echoes is developed,effectively enhancing the signal-to-clutter ratio(SCR in situations where the targets are submerged in the clutter environment.In addition,complex-valued network computations and gradient penalty terms are introduced,proposing a method based on complex-valued Wasserstein generative adversarial network for nonhomogeneous clutter suppression,further improving the model’s training stability and generalization capability.(2)A nonhomogeneous clutter suppression method based on dynamic encoder decoder augmentation network is proposed.A ResUblock structure and a contrast information augmentation module are designed to extract abundant contrast information between targets and clutters.A result consistency loss function addresses the issue that the model structure cannot dynamically adjust to tasks,effectively enhancing the SCR of weak targets in the complex clutter environment with lower-computational-cost.(3)A target detection method for range-profile features based on complex-valued UNet is proposed.Utilizing an up-down sampling structure with skip connections to merge complex features of different scales,and designing a constant false alarm rate controller,the single-pulse target detection performance under low SCR conditions is enhanced.Further,a range-profile and range-Doppler spectrum feature fusion target detection method based on complex-valued convolutional neural networks is proposed for multi-pulse signals.This method extracts the inter-pulse sequential correlations of echo range-profile and range-Doppler spectrum,and a feature fusion module with adaptive convolution weight learning is designed to achieve target detection in nonhomogeneous clutter environments.Additionally,by integrating polarization information,a multi-polarization feature fusion target detection method based on complex-valued long short-term memory networks is proposed,demonstrating the advantages of combining multiple polarization features in target detection.(4)A self-evolving target detection architecture based on incremental learning is proposed,which can continuously adapt to time-varying clutter environments.Through a three-level weighted evaluation system of the model,the performance of radar detection models is assessed in real-time,guiding the self-evolution direction of the model.Simultaneously,a sample selection method based on the nearest features mean is proposed,enabling rapid self-evolution of the detection model under few-shot conditions.The above methods have been validated through experiments with publicly available datasets,achieving stable detection of targets in nonhomogeneous time-varying clutter environments,laying a theoretical and technical foundation for advancing radar detection technology in complex environments. |
