PlakaTrainer

A pipeline for building Turkish license plate character training data from raw images.

Pipeline Overview

01_detect_plates.py  →  annotate_chars.py  →  training data (output/)
     (ONNX)               (Flask web UI)

1. Detect — YOLOv26 ONNX model finds plates, crops and deskews them into deskewed/
2. Annotate — Flask web annotator auto-segments characters per plate and lets you label each one
3. Output — Per-character labeled crops saved to output/ ready for CNN training

Files

File	Purpose
01_detect_plates.py	ONNX plate detection, crop + deskew → deskewed/
annotate_chars.py	Flask web annotator for character labeling
annotater.py	Earlier annotation utility
02_extract_digits.py	Legacy: Tesseract-based character extraction
03_review_digits.py	Legacy: OpenCV review/reclassify tool
kareplaka.onnx	Trained YOLOv26 plate detector
annotations.csv	Saved character annotations
debug_seg.py	Visual pipeline debugger (saves step images to debug_seg/)

Requirements

python3 -m venv ../venv
source ../venv/bin/activate
pip install -r requirements.txt

Python packages (requirements.txt):

• onnxruntime
• opencv-python
• flask
• numpy

Step 1 — Detect & Deskew Plates

Runs the ONNX detector on a folder of images, crops each detected plate, corrects tilt, and saves to deskewed/:

python 01_detect_plates.py --input-dir /path/to/images --output-dir .

Options:

Flag	Default	Description
--input-dir	required	Folder of source images
--model-path	kareplaka.onnx	ONNX model file
--output-dir	.	Parent of deskewed/ output folder
--confidence-threshold	0.25	Detection confidence cutoff
--max-angle	15.0	Max deskew correction angle (degrees)

Output: deskewed/*.jpg — one file per detected plate.

Step 2 — Annotate Characters

Launches a Flask web UI to label characters in each deskewed plate:

python annotate_chars.py --input ./deskewed/ --output ./output

Options:

Flag	Default	Description
--input	./deskewed	Folder of deskewed plate images
--output	./output	Destination for labeled character crops
--port	5000	HTTP port

Open http://localhost:5000 in a browser.

Annotator Features

• Auto-segments characters using a histogram + contour pipeline
• Click & drag on the plate image to add missing bounding boxes
• Click a box thumbnail to select it, then type the character label
• Enter — save & next plate
• ← / → — prev / next plate
• Del — remove selected box
• n — skip plate

Character Segmentation Pipeline

1. Flood-fill inner rect — heavy Gaussian blur dissolves characters; connected-component analysis with border-touching exclusion isolates the plate body blob
2. Binarise crop — CLAHE + Otsu; polarity determined from the binary itself so dark-frame plates don't confuse it
3. Contour pass — finds individual character blobs; wide merged blobs are histogram-split
4. Histogram fallback — column-sum histogram gap detection at progressively relaxed thresholds

Typical Workflow

# 1. Detect plates and deskew
python 01_detect_plates.py --input-dir /media/ce/data/plates --output-dir .

# 2. Annotate characters in the web UI
python annotate_chars.py --input ./deskewed/ --output ./output
# → open http://localhost:5000

Debugging Segmentation

debug_seg.py runs the full segmentation pipeline on a single image and saves 15 intermediate images to debug_seg/:

python debug_seg.py deskewed/03BDT873-20190615-132835.jpg

Saved images: 00_original → 01_gray → 02_heavy_blur → 03_thresh_raw → 04_thresh_polarity → 05_filled → 06_inner_contours → 07_inner_rect_raw → 08_inner_rect_inset → 09_crop_gray → 10_crop_clahe → 11_binary_crop → 12_col_histogram → 13_char_contours → 14_final_boxes