A Dual-use autoencoder for fingerprint enhancement and feature extraction

Fingerprint biometrics underpin many systems that require person recognition and yet live prints often fail to match templates due to poor image quality. This degradation can arise from acquisition noise, skin condition, pressure inconsis-tencies and other factors. Studying these failures requires access to prints that contain such noise, but fingerprint datasets are too sensitive to be publicly shared. We propose an autoencoder that can help with extracting this useful information from a dataset of fingerprints, without revealing identity-related ridge information that can be directly traced back to an individual. The autoencoder splits each fingerprint image into structure and style latent codes. The structure code holds identity-bearing ridge pattern information and the style code captures non-identity related noise and damage patterns. From the structure code alone, the model reconstructs a pristine version of the input, effectively enhancing poor-quality prints. Combining both the structure and style codes recreates the original noisy image. Agencies that collect fingerprints can thus publish only the style codes, so others can blend realistic noise into their own datasets for benchmarking, error analysis and for training neural network models. Crucially, this sharing of information enables researchers and developers to analyze and close demographic performance gaps, ensuring the resulting systems work more equitably for populations whose fingerprints are disproportionately affected by skin conditions, occupational wear, age-related changes and other challenging patterns of perturbation. They can also use the structure-only reconstruction to clean incoming prints and reduce matching errors.