IDTrust: Deep Identity Document Quality Detection with Bandpass Filtering
Core Concepts
IDTrust introduces a deep-learning framework for assessing the quality of IDs, enhancing dataset applicability and eliminating the need for original document patterns.
Abstract
IDTrust is a system developed to enhance the quality of identification documents by utilizing a deep learning-based approach. It aims to effectively detect and differentiate ID quality using bandpass filtering. The system offers significant improvements in dataset applicability by eliminating the need for relying on original document patterns for quality checks. By conducting experiments on datasets like MIDV-2020 and L3i-ID, optimal parameters were identified, significantly improving discrimination performance between original and scanned ID documents. The proposed models, DeepQD and GuidedDeepQD, outperform existing methods in accurately discerning between original and scanned IDs across different countries. They achieve near-perfect accuracy, F1-score, and AUC values, showcasing robust discrimination capabilities.
IDTrust
Stats
MIDV-2020 dataset includes 1000 video clips, 2000 scanned images, and 1000 photos of unique dummy IDs.
L3i-ID dataset consists of 17 types of original French IDs.
CheckScan model achieved varying levels of accuracy with higher precision settings yielding better results.
Quotes
"IDTrust eliminates the need for original document patterns and pre-processing steps."
"GuidedDeepQD consistently outperforms DeepQD on the L3i-ID dataset."
"The proposed models achieve perfect accuracy on the MIDV-2020 dataset."