Target Classification in Conventional Radar Using Cyclic Spectrum and Enhanced VGGNet

To overcome the limitations of conventional radar target recognition methods, this paper proposes a novel approach that integrates cyclic spectrum slicing with deep learning techniques. First, the cyclic spectrum of the radar signal is computed using the time smoothing method, enabling a theoretical analysis of the distinguishing features present in the target’s cyclic spectrum slice. The resulting slice spectrum is then fed into an improved Deep Convolutional Generative Adversarial Network (DCGAN) model, which employs a Wasserstein GAN with Gradient Penalty (WGAN-GP) loss function for data augmentation. Subsequently, an enhanced VGGNet architecture—incorporating Global Average Pooling and Parametric Rectified Linear Unit (PReLU) activation functions—is utilized to automatically extract features that capture the target’s cyclic stability. Experimental results on aircraft target recognition demonstrate that the proposed method achieves outstanding average accuracies of 98.46 % and 98.40 % for different cyclic spectrum slices, significantly outperforming traditional methods by more than 3 %. The findings highlight the exceptional capability of the cyclic spectrum in revealing the intrinsic properties of target signals, along with its strong suppression of noise and clutter. Moreover, the integration of deep learning techniques has substantially improved target recognition accuracy, offering an effective solution for automatic target recognition under conditions of limited samples and low signal-to-noise ratios.