Latent Fingerprint Enhancement and Segmentation Through Advanced Deep-Learning Techniques
Poornima E GundgurtiDepartment of Computer Science, Central University of Karnataka, Kadaganchi, Kalaburagi, Karnataka and Affiliated to Visvesvaraya Technological University, Belagavi, Karnataka, India. poornimae25@gmail.com0009-0009-1171-546X
Dr. Shrinivasrao B KulkarniDepartment of Computer Science and Engineering, SDM College of Engineering and Technology, Dharwad, Affiliated to Visvesvaraya Technological University, Belagavi, Karnataka, India. sbkulkarni_in@yahoo.com0000-0001-5576-5076
Keywords: Latent Overlap Fingerprint, Mask RCNN, Edge Directional Total Variation Model (EDTV), Normalization, Segmentation.
Abstract
Fingerprint identification in criminal investigations and biometric systems faces challenges because of poor image quality at crime scenes. This work addresses the complex task of latent fingerprint enhancement and segmentation through a Modified Deep Learning Model. The initial step involves normalisation to predefine mean and variance, mitigating grey-level volatility caused by ridges and valleys. Subsequently, a2 Edge Directional Total Variation model-based adaptable de-noise efficiently removes structured noise, enhancing latent fingerprint images. For segmentation, the Modified Mask R-CNN is proposed to identify critical features of overlapping latent fingerprints by assigning additional weight to neighboring borders, enhancing separation. The research introduces a strategy called Atrous-based Modified Mask RCNN with cascaded Atrous II-blocks, residual learning, and Instance Normalization, which proves accurate in improving and segmenting latent fingerprints. The result reveals the proposed technique utilises the Tsinghua Latent Overlapped Fingerprint Database for analysis, which attains a high accuracy of 0.99 compared with existing models.