LCO-BNN: A Lightweight Collaborative Optimization Binary Neural Network for Radar Jamming Recognition

Advancements in electronic technology have made the battlefield electromagnetic environment increasingly complex, posing challenges for radar systems in accurately detecting targets and countering jamming. Especially lightweight platforms such as unmanned aerial vehicles and individual soldier devices have limited computational resources, and there is an urgent need for jamming recognition models with low power consumption and strong real-time performance. However, existing deep learning-based methods suffer from high computational costs, large model sizes in the algorithms, and poor adaptability to composite jamming scenarios. To address these limitations, this paper proposes a Lightweight collaborative optimization binary neural network (LCO-BNN) for radar jamming recognition. Firstly, the LCO-BNN integrates a two-stream input that combines Cross-ambiguity function (CAF) contours and Range-doppler (RD) spectrograms to extract complementary spatiotemporal features. Secondly, although Binary neural networks (BNNs) with 1-bit weights and activations significantly reduce computational costs, their severe information degradation hinders deployment. To mitigate this, we propose two mechanisms: Variance-aligned adaptive scaling (VAS) in forward propagation, which enables the binary weights to approximate the statistical distribution of the original weights and effectively preserve a greater amount of information, and Gradient-adaptive coordination (GAC) in backward propagation to guarantee stable gradient updates. Furthermore, the network employs Binary depthwise separable Convolutions (Bi-DS Conv) and Real-valued retention Skip (RRS), which efficiently balance efficiency and accuracy while lowering computational complexity and information loss. Experimental results demonstrate that the proposed LCO-BNN achieves a recognition accuracy of 99.33% under mixed JNR conditions. Compared with other full-precision models such as 2D-CNN, it significantly reduces the number of parameters to 4.30 Mbit and computational 1.34×106FLOPs, minimizing resource overhead by orders of magnitude. Furthermore, it outperforms existing binary methods with an accuracy improvement of at least 5.92%, providing a robust and low-power solution for real-time jamming recognition in resource-limited battlefield scenarios.