Decimal place separable complex convolutional neural network for wideband beamforming

With the rapid development of synthetic aperture radar and increasing demand for remote sensing, wideband beamforming technology has been a hot spot. Recently, deep learning from computer science has given a hint for the next generation of beamforming. Many neural network-based beamformers have been reported. However, those methods show disadvantages on wideband signals. Due to the precision limitations of current computing devices, neural networks may encounter precision errors when generating extremely small numerical values. However, the performance of beamforming is still highly sensitive to small numerical values. Existing neural network-based methods produce decreased performance due to those errors. To enhance the precision of generation and achieve a better trade-off between performance and efficiency, we propose a generation mechanism with high precision at the framework level. In this paper, the decimal place separable complex convolutional neural network (DSCCNN) is proposed for wideband beamforming. Firstly, we apply different networks to handle distinct decimal places contributing to a mixture-of-experts framework, which can increase the fitting precision. Then, multilayer perceptrons are used to enhance the learning capabilities of the proposed network’s backbone referring to current popular computer vision network architectures. Last, an improved attention module is proposed to better process the different parts of complex-valued feature maps based on the squeeze-and-excitation module. Simulation experiments show the proposed beamforming method has excellent performance in anti-jamming. The computational complexity of the proposed method is low, which is beneficial for potential engineering applications. In addition, the proposed network can be trained within a very short time.