Article directory
2018
Latent Fingerprint Recognition: Role of Texture Template
code: https://github.com/prip-lab/MSU-LatentAFIS
Abstract : A texture template method consisting of a virtual set of details is proposed to improve the overall latent fingerprinting accuracy. To compensate for the lack of a sufficient amount of detail in low-quality potential prints, we generated a set of dummy details. However, due to the large number of regularly placed virtual details, texture-based template matching is more computationally demanding than matching true detail templates. To improve the accuracy and efficiency of texture template matching, we investigate: i) training Convolutional Neural Networks (ConvNets) on raw and augmented fingerprint patches to improve descriptor specificity with each virtual detail, 2) small patches between virtual details and The Fast Combination architecture speeds up descriptor extraction, iii) reduces descriptor length, iv) a modified hierarchical graph matching strategy improves matching speed, and v) extracts multiple texture templates to improve performance. Experiments on the NISTSD27 latent database show that the above strategy can increase the matching speed from 11ms (24 threads) (between latent printing and reference printing) to 7.7ms (single thread), while improving the rank 1 accuracy by 8.9% over the 10K library .
Contribution of the paper:
- Reduced the average recognition time between latent texture templates and rolling texture templates from 11ms (24 threads) to 7.7ms (single thread);
- The rank 1 recognition rate of the 10K gallery texture template is improved by 8.9% (from 59.3% to 68.2%).
- By fusing the proposed three texture templates with the three templates in [4] (from 75.6% to 78.3%), the rank 1 recognition rate in [4] is improved by 2.7%. This means that out of the 258 latencies of NISTSD27, the improvement in texture stencils will be 7 extra latencies in rank 1.
