A³R-Net: adaptive attention aggregation residual network for sparse DOA estimation

In this paper, a unified deep learning framework is developed for high-precision direction-of-arrival (DOA) estimation. Unlike previous methods that divide the real and imaginary parts of complex-valued sparse problem into two separate input channels, a real-valued transformation is adopted to encode the correlation between them. Then, a novel adaptive attention aggregation residual network (A³R-Net) is designed to overcome the challenges in the case of low signal-to-noise ratios or small inter-signal angle separations. First, to alleviate the gradient disappearance and gradient explosion caused by network deepening, a residual learning strategy is introduced to construct a deep estimation network that learns the inverse mapping from the array measurement vector to the original spatial spectrum. Second, since the feature fusion method via simple summation in the shortcut connection ignores the inconsistency on the scale and semantic of features, an adaptive attention aggregation module (A³M) with adaptive channel context aggregators is proposed to capture multi-scale channel contexts and generate element-wise fusion weights. Finally, a dilated convolution with a broader receptive field is embedded into the channel context aggregator to learn wider local cross-channel association. Extensive simulation results demonstrate the superiority and robustness of the proposed method compared with other state-of-the-art methods.