Cycle-consistent generative adversarial neural networks based low quality fingerprint enhancement

Distortions such as dryness, wetness, blurriness, physical damages and presence of dots in fingerprints are a detriment to a good analysis of them. Even though fingerprint image enhancement is possible through physical solutions such as removing excess grace on the fingerprint or recapturing the fingerprint after some time, these solutions are usually not user-friendly and time consuming. In some cases, the enhancements may not be possible if the cause of the distortion is permanent. In this paper, we are proposing an unpaired image-to-image translation using cycle-consistent adversarial networks for translating images from distorted domain to undistorted domain, namely, dry to not-dry, wet to not-wet, dotted to not-dotted, damaged to not-damaged, blurred to not-blurred. We use a database of low quality fingerprint images containing 11541 samples with dryness, wetness, blurriness, damages and dotted distortions. The database has been prepared by real data from VISA application centres and have been provided for this research by GEYCE Biometrics. For the evaluation of the proposed enhancement technique, we use VGG16 based convolutional neural network to assess the percentage of enhanced fingerprint images which are labelled correctly as undistorted. The proposed quality enhancement technique has achieved the maximum quality improvement for wetness fingerprints in which 94% of the enhanced wet fingerprints were detected as undistorted.