Advanced Radar Signal Processing Using Adaptive Threshold Convolution Neural Network

Abstract

Radar Automatic Target Recognition (RATR) is critical for surveillance in complex electromagnetic environments. Traditional methods struggle with interference suppression, while existing deep learning approaches lack adaptability to dynamic Signal-to-Interference Ratios (SIR). This paper proposes an Adaptive Threshold Convolutional Neural Network (ATCNN) featuring three innovations: a convolution unit that dynamically adjusts activation thresholds using real-time SIR to suppress noise and enhance feature extraction; a multi-scale framework with varied kernel sizes to capture global and local patterns in High-Resolution Range Profiles; and a channel attention mechanism fused with residual connections to prioritize salient features while preserving data integrity. Evaluations on measured High-Resolution Range Profile (HRRP) datasets show ATCNN’s superior accuracy over existing methods. It maintains stable performance under diverse interference conditions across SIR levels, outperforming baseline models. Ablation studies confirm each module’s necessity, with performance dropping significantly when core components are removed. The framework’s environmental adaptability and balanced feature extraction advance target recognition in real-world scenarios, offering scalable solutions for autonomous systems application.