Residual Units and Attention Gates-Based Asymmetric Encoder-Decoder Network for Sea Clutter Suppression
To address the challenge of weak target detection in nonhomogeneous sea clutter environments,this paper proposes an asymmetric encoder-decoder network(AED-Net)for sea clutter suppression based on complex-valued residual units and attention gates.The network takes the complex-valued signal generated by radar echoes passing through the matched filter as input.First,it replaces conventional convolutional units with complex-valued residual units to extract fea-tures of weak targets and sea clutter,enhancing the network's capability of feature extraction while avoiding feature degra-dation.Then,attention gate modules are employed to selectively propagate the feature information extracted by each mod-ule in the encoding path to the corresponding modules in the decoding path.Finally,it yields the complex signal after sea clutter suppression as output.Due to the capability of independently selecting the input and output dimensions of each atten-tion gate according to the network structure,the proposed network has an asymmetric encoder-decoder structure.Compared to typical symmetric encoder-decoder network,UNet,the introduction of complex-valued residual units and attention gates significantly reduces the redundancy of feature information,enhances feature extraction and transmission,and thus im-proves the sea clutter suppression performance.Meanwhile,complex-valued residual units have much smaller parameter size than convolutional units and the introduced attention gates greatly reduce the number of units in the decoding path,re-sulting in a significant reduction of total network parameters.Experimental results based on real sea clutter data demonstrate that compared to the complex value-UNet(CV-UNet)network,AED-Net achieves an average improvement of 9 dB in the output signal-to-clutter ratio(SCR)and can effectively operate at a minimum SCR reduction of 3 dB.Moreover,the number of parameters and the computational cost are reduced by 57.8％and 50％,respectively.
sea clutter suppressionencoder-decoder networkresidual unitattention gatecomplex-valued signal
万方数据
维普
