Universal Blind-Denoising Method of Radar Spectrograms for Unknown Noise Distribution

Human target detection and perception technology has numerous potential applications, including autonomous driving, biometric recognition, and anomaly detection. The distinctive advantages of radar sensors include their ability to function normally in a wide range of lighting and weather conditions, their low energy consumption, and their capacity to protect personal privacy. In recent years, they have become an important research topic in human target detection and perception technology. Due to the interference of channel clutter and environmental noise in the received radar echo signal in practical applications, it is necessary to design denoising models or methods to suppress clutter and noise in the radar spectrograms in order to prevent its impact on subsequent human target perception and detection tasks. Due to the high cost of obtaining measured radar spectrograms, we propose using simulated radar spectrograms as training data and measured radar spectrograms as testing data to develop a universal blind-denoising model. We design data augmentation methods and optimization strategies based on the characteristics of the measured radar echo signals to improve the model’s generalization performance on out-of-distribution data and solve the problem of domain shift. In addition, we evaluate the denoising effect of the proposed model using six objective image quality evaluation metrics. The experimental results demonstrate that the proposed denoising model outperforms all comparison methods in nearly all metrics, proving its superiority in solving the problem of universal blind denoising for radar spectrograms.