Abstract:
Considering the challenges faced by existing algorithms in extracting highly discriminative domain-shared features from cross-domain footprint images, a cross-domain footprint image retrieval method based on feature partitioning and domain fusion is proposed. Firstly, a cross-domain footprint image dataset containing 200 individuals is constructed, and the characteristics of footprint images in two domains are analyzed. Secondly, a feature-level partitioning method is introduced to obtain more discriminative local features. Lastly, a cross-domain feature fusion method is employed to extract domain-shared information and design an equilibrium loss function to optimize the fused features. The proposed method performs well on the cross-domain footprint image dataset, achieving Rank-1 of 91.37% and 84.50% under optical retrieval pressure and pressure retrieval optical settings, respectively.