Detecting the National Identification Cards

English | Türkçe

You can use this project to extract information DOB (name, surname, date of birth, etc.) on the identity card. To do this, I'm broke down the problem into sub-problems as below:

• [this project] Identify Regions of Interest (ROI) containing the required information with deep learning
• [this project] Crop the regions identified above
• OCR on the identified region of interest

This project can do object detection + object classification + multiple object detection all at the same time.

Use case diagram

Sample id cards

Sample 1	Sample 2	Sample 3

Modernization Notes (2025)

This project has been refreshed after 7+ years. It now uses TensorFlow 2 (SavedModel) with optional OCR (EasyOCR).

• Removed TensorFlow 1 graph code and TF Object Detection API dependency. No more object_detection.* imports.
• Loads TF2 SavedModel from model/saved_model/; automatically resolves and calls the serving_default signature.
• id_card_detection_image.py CLI flags:
  • --image: Image path (absolute/relative)
  • --min_score: Score threshold (0–1, default 0.60)
  • --ocr: Run OCR (EasyOCR) on the cropped ROI
• Lightweight label map parser added; reads ID→name mapping from data/labelmap.pbtxt.
• Visualization via OpenCV; top-scoring box in green, others red. Cropped ROI is saved as output_cropped.png.

Note: SavedModel output keys can vary between models. Defaults expect detection_boxes, detection_scores, detection_classes. Adjust easily if different.

Setup and Run (Recommended TF2 + OCR)

Install either a single platform-specific requirements file (includes TensorFlow), or install TensorFlow separately plus requirements-modern.txt. YOLO is optional and can be installed with ultralytics.

Option A) One-shot install (recommended)

• Apple Silicon (macOS):

pip3 install -r requirements-macos-apple.txt

• Intel macOS / Linux / Windows (CPU):

pip3 install -r requirements-cpu.txt

Option B) Split install (advanced)

Install TensorFlow first, then the rest from requirements-modern.txt.

1) Install TensorFlow

• macOS (Apple Silicon, M-series):

pip3 install tensorflow-macos==2.16.1 tensorflow-metal==1.2.0

• macOS (Intel) or Linux (CPU):

pip3 install tensorflow==2.20.0

• Windows (CPU):

pip3 install tensorflow==2.17.1

Tip: On Python 3.12, keep pip up to date: python3 -m pip install --upgrade pip

2) Other dependencies

pip3 install -r requirements-modern.txt

3) Run

python3 id_card_detection_image.py --image /absolute/or/relative/path.jpg --ocr --min_score 0.6

• Windows example path: C:\\path\\to\\image.jpg
• Cropped ROI is written as output_cropped.png at project root.
• With --ocr, extracted text lines are printed to the terminal.

YOLO integration (optional)

Install YOLO backend:

pip3 install ultralytics

Run with YOLO instead of TF2 (image):

python3 id_card_detection_image.py --image /path/to/img.jpg --yolo_model yolov8n.pt --min_score 0.4 --ocr

Run with YOLO (camera) and enable OCR snapshot panel (keys 1–9 to OCR selected crop):

python3 id_card_detection_camera.py --yolo_model yolov8n.pt --min_score 0.4 --ocr

Camera window controls: q quit, p pause/resume, s stop camera, b start camera, 1–9 OCR selected snapshot.

Command-line Arguments

Image script (id_card_detection_image.py):

• --image (string): Absolute or relative path to the image file. If omitted, defaults to test_images/image1.png.
• --min_score (float, default 0.60): Minimum confidence score threshold in [0,1] to visualize and crop detections. Lower it (e.g., 0.3–0.5) to see more candidates.
• --ocr (flag): If provided, runs EasyOCR on the cropped ROI and prints extracted text lines to the terminal.
• --yolo_model (string, optional): If provided, switches detector backend to YOLO. Accepts a local .pt path or a model name like yolov8n.pt. If omitted, TF2 SavedModel in model/saved_model/ is used.

Camera script (id_card_detection_camera.py):

• --camera (int, default 0): Video device index. Try 1 or 2 if you have multiple cameras.
• --min_score (float, default 0.50): Minimum confidence score threshold for drawing detections and listing snapshots.
• --yolo_model (string, optional): YOLO model path/name, same behavior as in the image script.
• --ocr (flag): Enables the OCR workflow on detected snapshots. When enabled:
  • The right panel shows up to 9 cropped detections sorted by score.
  • Press number keys 1–9 to run OCR on the corresponding crop; results are shown under “OCR Result” in the right panel and printed to terminal.

Window hotkeys (camera):

• q: Quit.
• p: Pause/Resume the live feed (freezes on the last frame when paused).
• s: Stop (release) the camera device.
• b: Start/restart the camera device.
• 1–9: Run OCR on the Nth snapshot in the right panel (only if --ocr was provided).

Outputs:

• Image script: draws boxes on the image window and saves the best ROI as output_cropped.png (project root). OCR text lines print to terminal.
• Camera script: draws boxes over the live feed; right panel lists thumbnails and shows OCR results if --ocr is enabled.

Alternative: Legacy TF1 Flow

Requires Python 3.7 and TensorFlow 1.15. Hard to set up on modern systems.

pip3 install -r requirements.txt
python3 id_card_detection_image.py

Warning: TF1 deps (esp. on macOS/ARM) are incompatible with modern Python. Prefer Docker or a dedicated Python 3.7 env if needed.

Platform Notes

• macOS Apple Silicon: tensorflow-macos + tensorflow-metal required. First run may build font caches.
• macOS Intel / Linux: CPU tensorflow is sufficient; GPU not required.
• Windows: pip install tensorflow==2.17.1 (CPU). Ensure Visual C++ Redistributable is installed if needed.

Examples

• Recommended flow (macOS Apple Silicon):

pip3 install tensorflow-macos==2.16.1 tensorflow-metal==1.2.0
pip3 install -r requirements-modern.txt
python3 id_card_detection_image.py --image /Users/you/Downloads/test_data.jpg --ocr --min_score 0.6

• Recommended flow (Linux / macOS Intel):

pip3 install tensorflow==2.20.0
pip3 install -r requirements-modern.txt
python3 id_card_detection_image.py --image ./test_images/image1.png --min_score 0.6

Result