An On-Premise Deep-Learning Pipeline for Robust Machine-Readable-Zone Extraction from Real-World Identity-Document Images

Automated reading of the Machine-Readable Zone (MRZ) of identity documents is a central step in digital identity verification, but robust extraction from real-world smartphone images remains difficult because of blur, glare, perspective distortion, and partial occlusion. Production-grade solutions are typically commercial, closed-source, and operate as black boxes, which is undesirable for privacy-sensitive on-premise deploy- ments. This paper presents and evaluates a fully on-premise pipeline that converts a complete document image into an MRZ transcription through four stages: MRZ detection, geometric rectification, line splitting, and MRZ-adapted optical charac- ter recognition (OCR). For detection, a lightweight literature- based HED-MRZ baseline, a U-Net, and two vision-foundation- model backbones (DINOv2-S and DINOv3-L) are compared as binary segmentation models. For recognition, generic PaddleOCR systems and a commercial OCR service are compared against an SVTRv2-based connectionist temporal classification (CTC) recognizer that is pretrained on synthetically generated MRZ lines and fine-tuned on real MRZ crops. Models are evaluated on the public MIDV-500, MIDV-2020, and MIDV-LAIT datasets and on a private target-domain dataset, with paired statistical testing. U-Net, DINOv2-S, and DINOv3-L clearly outperform the HED-MRZ baseline, and MRZ-specific OCR adaptation substantially improves recognition over generic OCR. On the private target domain the end-to-end pipeline reaches 96.4% full- MRZ accuracy with a character error rate of 0.07%. The results indicate that accurate on-premise MRZ extraction is feasible without relying on closed commercial OCR services. Index Terms—Machine-Readable Zone, optical character recognition, identity documents, semantic segmentation, vision transformers, document image analysis.