An End-to-End Deep Learning Architecture for Information Extraction from Turkish Identity Cards

Authors : Ömer Can Eskicioğlu, Ali Hakan Işik

Pages : 81-86

Doi:10.54569/aair.1845016

View : 97 | Download : 80

Publication Date : 2025-12-23

Article Type : Research Paper

Abstract :With the acceleration of digital transformation in the service sector, remote customer acquisition and identity verification processes have become the cornerstone of secure ecosystems. Particularly in internet-based services, image distortions, perspective errors, and variable lighting conditions encountered during the transfer of physical documents to the digital environment are the most significant factors complicating data extraction. In this study, a deep learning-based end-to-end architecture is proposed that enables fast, secure, and high-accuracy information extraction from Turkish Republic Identity Cards. In the proposed system, while CURL is used to enhance image quality, a YOLOv8m-based instance segmentation model is preferred for detecting the boundaries of the card. For the determination of card orientation and perspective correction, a novel hybrid approach has been developed that analyzes the cosine distance between face biometrics obtained via RetinaFace and the segmentation mask. This structure, in which PAN for text detection and Transformer-based TrOCR models for character recognition are integrated, was tested on a unique dataset augmented with the CLoDSA library. Experimental results indicated that the YOLOv8m model exhibited success in card detection with a 99.5% mAP score. Our proposed model demonstrates that it offers an efficient solution for digital identity verification processes with an overall accuracy rate of 92.6%.
Keywords : Dijital Kimlik Doğrulama, YOLOv8, RetinaFace, Perspektif Düzeltmesi, TrOCR, Derin Öğrenme

ORIGINAL ARTICLE URL