2022–2023
Recognition & OCR
Real-time face-mask compliance, face recognition, and Arabic ID reading — the computer-vision and OCR work that taught me to pull structure out of messy pixels.
- YOLO
- CNN
- Siamese networks
- Tesseract
- FaceVGG
- LoRA/QLoRA
- ChromaDB
- FAISS
The problem
Vision problems in the wild are never clean. A CCTV frame in a crowded corridor has to answer 'is this person masked?' in real time, at a glance, under bad lighting. A face-recognition gate has to tell two strangers apart from a single reference photo. And an Egyptian national ID mixes cursive Arabic, printed fields, and a face crop that off-the-shelf OCR simply chokes on. Each one is the same underlying job: turn noisy pixels into a decision you can act on.
The approach
- 01
Real-time mask detection with YOLO + CNN
During COVID-19 I built a mask-compliance detector: YOLO locates every face in the frame and a CNN classifier calls masked vs. unmasked, fast enough to run live on a video feed rather than batch-processing stills.
- 02
One-shot face recognition with Siamese networks
Instead of training a classifier per person, a Siamese network learns a similarity metric — so a single enrolment photo is enough to verify identity at the gate, and adding a new person needs no retraining.
- 03
Egyptian-ID Arabic OCR pipeline
I combined YOLO to detect and crop each field, Tesseract tuned for Arabic to read the text, and FaceVGG to extract and verify the ID photo — turning a whole card into structured, queryable fields.
- 04
LLMs on a single GPU + RAG
To make the extracted text useful, I ran LLMs on one GPU with LoRA/QLoRA fine-tuning, and grounded them in a retrieval layer over ChromaDB and FAISS so answers came from real documents, not guesses.
The result
- Real-time live mask-compliance detection on video, not batch
- ~95% conf. mask-detection confidence (illustrative)
- One-shot face verification from a single reference photo
- ~90% acc. Arabic-ID field OCR (illustrative)
This is where the thread of my work starts: everything from clean pixels to a usable decision. The same instinct later drove the peer-reviewed LCBVAE research on dotted Arabic dates — decode a clean signal out of noise, whether the output is a mask alert, an identity, or a parsed ID.
Back to portfolio