An End-to-End Computer Vision System for Structured Information Extraction from Turkish ID Card Images

Download PDF
Open Webpage

Seyf Kazamel, Umut Kocasari, Ali Alici

Published: 2025, Last Modified: 28 Feb 2026ICCSA (2) 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: This paper introduces an end-to-end computer vision pipeline for automatically extracting key personal information from Turkish ID card images. The system addresses real-world challenges such as invalid images, incorrect card orientation, and older ID card formats by integrating advanced methods for card detection, card format classification, orientation correction, text region segmentation, and optical character recognition (OCR). Specifically, we employ YOLOv5 for ID card detection, CLIP to validate that the image is an official Turkish ID and to distinguish between old and new ID formats, CRAFT and U-Net to segment relevant textual regions, and Tesseract for final OCR. By interfacing with APIs from the Turkish Population Registry, our system performs near real-time ID validation while significantly reducing manual effort. Experimental results indicate over 90% accuracy in identifying and classifying ID cards, with OCR word-level accuracy surpassing 96%, and an end-to-end card-level accuracy of 73% for the pipeline, all with execution times under five seconds. Deployed in production, the system saves manual labor hours and speeds up processing workflows. Our findings underscore the feasibility and business value of automated ID verification, highlighting how deep learning–driven ID processing can streamline operations, mitigate fraud, and enhance the user experience.