LatentPrintFormer: A Hybrid CNN-Transformer with Spatial Attention for Latent Fingerprint identification

The recent publication of 'LatentPrintFormer' marks a significant advancement in latent fingerprint identification, a field plagued by challenges such as low image quality and background noise. This innovative model integrates a CNN backbone (EfficientNet-B0) with a Transformer backbone (Swin Tiny), enabling it to extract both local and global features effectively. A spatial attention module further enhances the model's performance by emphasizing high-quality ridge regions while suppressing noise. The features are then projected into a unified 512-dimensional embedding, allowing for efficient matching through cosine similarity in a closed-set identification setting. Extensive experiments on publicly available datasets demonstrate that LatentPrintFormer consistently outperforms three state-of-the-art techniques, achieving higher identification rates across Rank-10. This breakthrough not only enhances the accuracy of fingerprint recognition but also holds significant implications for for…