Hybrid Attention Conditional Generative Adversarial Network Based Framework for Latent Fingerprint Enhancement

This paper presents a novel Hybrid Attention-Guided Conditional Generative Adversarial Network Framework (Hybrid cGAN) for latent fingerprint enhancement, which provides a solution for improving low-quality latent fingerprints widely encountered in latent fingerprint analysis, i.e., noise, blur and partial ridge. The proposed methodology utilizes the data from the IIITD Latent Fingerprint Dataset to carry out image preprocessing of the latent fingerprint, using contrast-limited adaptive Histogram Equalization (CLAHE) and data augmentation via synthetic degradations as a simulation of system and problems occurring in a real-world scenario. A multi-scale CNN backbone effectively extracts the hierarchical fingerprint features and is refined using a dual attention mechanism with channel and spatial attention modules. As a result, incorporating a dual attention mechanism distinctly improves the enhancement, concentrating on crucial ridge channels and spatial areas where enhancing fingerprint is paramount. Besides, a conditional Generative Adversarial Network (GCN) is incorporated to restore latent fingerprint images conditioned on attention-refined features. Further, steps are used to enhance the generative restoration with orientation, frequency, and perceptual losses that specialize to the fingerprint domain to improve structural integrity and visual realism.