Transfer Learning-Based Target Detection Under Compound Gaussian Clutter

Accurate target detection amid noise and clutter remains a fundamental challenge in radar signal processing, particularly for applications such as autonomous navigation and defense systems. The presence of noise and clutter significantly distorts the reflected radar signals, making reliable detection more difficult. Recent developments in deep learning-especially the use of convolutional neural networks and transfer learning-have shown potential in improving radar signal classification performance. This study investigates the application of ResNet architectures for detecting multiple targets, and evaluates their effectiveness against classical constant false alarm rate techniques and adaptive normalized matched filters under conditions of correlated non-Gaussian clutter and low signal-to-clutter-plus-noise ratios. By leveraging pre-trained models with strong nonlinear feature extraction capabilities, the proposed approach demonstrates superior detection probability, outperforming both traditional and current state-of-the-art methods in radar signal processing tasks.