Magnetic Anomaly Target Detection Using Neural Network With 2-D Multihead Attention

Magnetic anomaly target detection is a widely used technique to locate obscured ferromagnetic objects for both civilian and military purposes. To improve the detection performance at low signal-to-noise ratios, we propose a deep learning network based on 2D multihead attention in this paper. First, three types of samples are prepared, including pure magnetic noise, noise mixed with complete signals, and noise mixed with truncated signals of different percentages. Then, the 1D time sequences are converted as 2D time-frequency spectra using wavelet synchrosqueezed transform for better feature extractions. Afterwards, a network is trained by using the time-frequency spectra for classification of the samples. The network mainly consists of two 2D multihead attention modules and several convolutional layers. The first convolutional layer uses the time-frequency spectra of three orthogonal basis functions as the convolution kernels to extract the rough signal features. The parameters of the 2D multihead attention are obtained by convolution operation for fine feature extractions. Experimental results demonstrate that the proposed scheme has higher detection performances than existing approaches to some extent, especially at low signal-to-noise ratios. It also shows that training the network using both complete and truncated signal samples is essential to raise the detection rate and reduce the false alarms.