Published Paper Details:

The ability to reconstruct fingerprints is an important element in biometric security systems, as well as forensic investigations. The purpose of this research is to study whether a convolutional autoencoder, modelled off a U-net architecture, could be used to reconstruct fingerprints from quality degraded or incomplete fingerprints. The model was also trained with pre-processing of a publicly available fingerprint dataset to convert the data to greyscale, resize them to uniform dimensions and normalize them. The model implemented a hybrid loss of Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM) loss to promote an accurate pixel match as well as structural similarity. Reconstruction results showed some promise in reconstructing core ridge structures of a fingerprint and highlighted the potential for digitized images of convolutional autoencoders to reconstruct and apply to fingerprint image reconstruction tasks.