erdem-ocr: Turkish line OCR for microcontrollers

Newer version: inviolus/InviolusOCR (2026-10-31) is smaller and its int8 file reads 78% of held-out lines exactly instead of 33%.

A compact CNN + BiLSTM + CTC recognizer for one cropped line of printed Turkish text, built to eventually run on a microcontroller-class device. It ships as a float Keras model and a full-int8 TensorFlow Lite model (~650 KB).

Türkçe özet: Mikrodenetleyiciler için tasarlanmış, tek satır basılı Türkçe metni okuyan küçük bir OCR modeli (yaklaşık 631 bin parametre, int8 TFLite dosyası yaklaşık 650 KB). Ticari olmayan kullanım içindir (CC BY-NC 4.0).

Files

File	What it is
`erdem_ocr.keras`	Float recognition model, input `(batch, 32, 160, 1)`, output `(batch, 40, 90)` softmax
`erdem_ocr_int8.tflite`	Full-int8 export, static batch 1, input `[1, 32, 160, 1]` int8, output `[1, 40, 90]` int8
`alphabet.json`	Character set and blank index used to decode the output

Model

Input: one horizontal text line, grayscale, resized to 160×32 (width×height), scaled to [-1, 1] (pixel / 127.5 - 1).
Architecture: 4 conv + batch-norm + pool blocks, 2 × bidirectional LSTM (96 units), dense softmax. 631,226 parameters.
Output: 40 timesteps × 90 classes. Class 0 is padding (never predicted), classes 1..88 are the characters of alphabet.json in order, class 89 is the CTC blank.
Character set: digits, ASCII letters, ç ğ ı İ ö ş ü Ç Ğ Ö Ş Ü, space and .,:;!?-_/()'%. Up to 24 characters per line.
Decoding: greedy CTC (argmax per step, collapse repeats, drop blanks).
int8 quantization: input scale 0.00784, zero point -1; output scale 0.00390625, zero point -128.
Not included: text detection. The model expects a line that has already been cropped.

Usage

import json
import numpy as np
import tensorflow as tf

alphabet = json.load(open("alphabet.json", encoding="utf-8"))["alphabet"]
BLANK = len(alphabet) + 1

def preprocess(path):
    image = tf.io.decode_image(tf.io.read_file(path), channels=1, expand_animations=False)
    image = tf.image.resize(image, [32, 160])
    return (tf.cast(image, tf.float32) / 127.5 - 1.0).numpy()

def decode(ids):
    text, previous = [], None
    for i in ids:
        if i != previous and 0 < i < BLANK:
            text.append(alphabet[i - 1])
        previous = i
    return "".join(text)

model = tf.keras.models.load_model("erdem_ocr.keras", compile=False)
probabilities = model.predict(preprocess("line.png")[np.newaxis])
print(decode(probabilities[0].argmax(-1)))

Results

Exact-line accuracy (whole line correct) and character error rate (CER) of the float model, on data it was not trained on:

Test set	Samples	Exact line	CER
TS-TR official test split (real scene text)	2,078	45.33%	24.17%
Held-out synthetic + real + Zenodo lines	3,234	58.23%	12.61%
of which Zenodo (clean printed lines)	200	91.00%	1.61%
Small real-photo test set	66	19.70%	50.21%

int8 TFLite model: on the 3,993-row held-out split it reads 10.5% of lines exactly (CER 21.7%), against 55.7% for the float model. The quantization loss is large and unresolved, so treat the .tflite file as an experimental export.

Limitations

Printed text only; no handwriting, curved text or full pages.
Works best on clean, high-contrast lines; real scene text is still weak.
The int8 model loses most of the float model's accuracy (see above).
Text detection and line cropping must be done before this model runs.

Training data

Synthetic Turkish lines rendered from a 50k-word Turkish word list in several fonts, with blur, noise and perspective changes.
Turkish Scene Text Recognition (TS-TR), Serdar Yıldız, CC BY-NC 4.0.
Turkish OCR Text Image Dataset, Zenodo, CC BY 4.0.
A small number of the author's own photographs of printed text.

Training/validation splits are group-disjoint (no font, background or source group appears in both).

License

Because the training data includes TS-TR (CC BY-NC 4.0), this model is released under Creative Commons Attribution-NonCommercial 4.0:

You may use, share and adapt the model for non-commercial purposes.
You must credit this model and the datasets listed above.
Commercial use is not permitted.

The training code is MIT-licensed: ErdemWilkinson/InviolusOCR.

Version

Trained 2026-10-09. Later versions of this model score higher (65% exact on TS-TR) but were trained with a dataset whose license is not stated, so they are not published here.

Downloads last month: 40