Semi-Global Attention and Inter-Class Multi-constraint Enforced Open-Set Radar Emitter Individual Identification

This paper proposes an open-set radar emitter individual identification algorithm based on semi-global attention and inter-class multi-constraint to address the problem of inadequately adaptive feature representation and discrimination difficulty with out-of-distribution samples. By introducing semi-global attention mechanism combined with dual-path convolutional neural network, individual feature representation ability can be enhanced. The mechanism combines semi-global geometric similarity and learnable similarity. Through adaptively adjusting the attention score and retaining long-distance feature association, more accurate feature details and higher efficiency can be obtained. Furthermore, inter-class multi-constraint based loss function is presented in open-set scenario. By constructing multiple loss functions to constrain the distribution of known class samples in feature space, discriminated feature difference between known and unknown class samples can be achieved. In addition, class-based adaptive threshold selection strategy is proposed to achieve accurate rejection of open-set samples. Experimental results show that the proposed algorithm achieves identification accuracy of 97.55% in closed-set radar emitter data. In open-set scenarios, when maintaining a closed-set recognition rate of 91.78%, the rejection rate achieves 87.64%, which represents the proposed algorithm can effectively discriminate known individuals and unknown ones.