Theory-guided dual-stream axial interactive network for collaborative magnetic target localization and pattern recognition

Magnetic anomaly detection is widely used in unexploded ordnance identification, geological exploration, biomedicine, and other fields. Existing studies typically treat target localization and pattern recognition including shape and posture as separate tasks. Research on each task also faces two common challenges: analytical inversion methods based on physical models face difficulties in the presence of environmental interference, while purely data-driven models rarely incorporate classical theoretical guidance. To address these issues, this paper proposes a new method that integrates theoretical analysis with deep learning. First, through theoretical analysis, we reveal the distinct response mechanisms of the magnetic gradient tensor and magnetic tensor contraction to target position and pattern recognition, both of which are closely related to the target’s principal axis directions. Guided by this structured theoretical insight, a dual-stream axial interactive network is designed. The network employs a dual-stream architecture to process the magnetic gradient tensor and magnetic tensor contraction separately. By independently encoding and initially extracting features from data along each principal axis, the network effectively avoids axial confusion. Furthermore, a bidirectional cross-attention interactive network is introduced to establish a dynamic, content-aware information exchange channel between the features of the two sub-networks, enabling the network to learn the nonlinear coupling relationships between the two types of quantities and ultimately fuse them into a discriminative joint feature representation. Based on this joint feature representation, the model can simultaneously output three-dimensional target position and pattern recognition information. Experimental results on real-world data demonstrate the competitive performance of the proposed method, and further suggest that integrating domain knowledge with deep learning holds potential for practical applications.