hermitkk/alphabet-classifier

Handwritten Character Classifier (EMNIST, MobileNetV2)

Two MobileNetV2-based ONNX models for classifying grayscale images of handwritten characters. Both models predict case-insensitively — they are trained on both upper and lowercase handwriting but always output a single canonical label per letter.

Model	Classes	Val Accuracy	ONNX
Alphanumeric	Digits 0–9 + Letters A–Z + blank (37 total)	91.42%	outputs/exports/alphanumeric_model.onnx
Alphabet	Letters A–Z + blank (27 total)	96.18%	outputs/exports/alphabet_model.onnx

Use the alphabet model when your input is guaranteed to be a letter (higher accuracy, no digit/letter confusion). Use the alphanumeric model when the input may be a digit or a letter.

---

Model Details

Shared architecture

Property	Value
Backbone	MobileNetV2 (pretrained ImageNet)
Head	Linear(1280→256) → BatchNorm → ReLU → Dropout(0.3) → Linear(256→N)
Parameters	~2.56M
Input	96 × 96 grayscale (expanded to 3-channel internally)
Format	ONNX (opset 17)

Alphanumeric model — 37 classes

Property	Value
Classes	0–9 (indices 0–9), A–Z (indices 10–35), blank (index 36)
Dataset	EMNIST byclass — both upper and lowercase handwriting, labels folded to uppercase
Train samples	711,932
Val samples	118,323
Val accuracy	91.42%
Macro avg F1	0.917

Alphabet model — 27 classes

Property	Value
Classes	A–Z (indices 0–25), blank (index 26)
Dataset	EMNIST letters — upper and lowercase handwriting already merged at source
Train samples	138,800
Val samples	22,800
Val accuracy	96.18%
Macro avg F1	0.960

---

Training Configuration

Setting	Value
Optimizer	AdamW — backbone lr × 0.1, head lr 5e-4, weight_decay 1e-4
Scheduler	LinearLR warmup (5 epochs) → CosineAnnealingLR
Loss	CrossEntropyLoss with inverse-frequency class weights + label smoothing 0.10
Batch size	256
Max epochs	50
Early stopping	patience 10
Augmentation	RandomAffine, RandomPerspective, ColorJitter, GaussianBlur, RandomErasing
Mixed precision	AMP (CUDA only)
Blank class	Synthetic white images; hard-floored weight ≥ 3.0

---

Per-class Performance

Alphanumeric model

Digits

Class	Precision	Recall	F1
0	0.741	0.686	0.712
1	0.748	0.693	0.719
2	0.984	0.949	0.967
3	0.997	0.995	0.996
4	0.989	0.974	0.981
5	0.985	0.928	0.956
6	0.987	0.976	0.981
7	0.994	0.997	0.995
8	0.994	0.988	0.991
9	0.953	0.962	0.958

Letters

Class	Precision	Recall	F1
A	0.977	0.969	0.973
B	0.914	0.971	0.942
C	0.966	0.985	0.975
D	0.961	0.976	0.969
E	0.992	0.989	0.990
F	0.985	0.983	0.984
G	0.818	0.821	0.820
H	0.967	0.982	0.974
I	0.545	0.684	0.607
J	0.930	0.950	0.940
K	0.987	0.993	0.990
L	0.573	0.537	0.554
M	0.992	0.998	0.995
N	0.985	0.984	0.984
O	0.648	0.696	0.671
P	0.986	0.994	0.990
Q	0.783	0.761	0.772
R	0.986	0.982	0.984
S	0.913	0.978	0.944
T	0.989	0.989	0.989
U	0.969	0.955	0.961
V	0.912	0.956	0.934
W	0.988	0.997	0.992
X	0.976	0.990	0.983
Y	0.896	0.946	0.920
Z	0.762	0.921	0.834

Blank

Class	Precision	Recall	F1
blank	1.000	1.000	1.000

Hardest cases are I (F1=0.607) and L (F1=0.554), both confused with digit 1; and O (F1=0.671), confused with digit 0. These are inherent digit–letter ambiguities in alphanumeric OCR.

---

Alphabet model

Class	Precision	Recall	F1
A	0.959	0.976	0.968
B	0.996	0.986	0.991
C	0.984	0.979	0.981
D	0.976	0.974	0.975
E	0.984	0.986	0.985
F	0.994	0.976	0.985
G	0.927	0.874	0.900
H	0.976	0.978	0.977
I	0.760	0.761	0.761
J	0.974	0.965	0.969
K	0.996	0.995	0.996
L	0.765	0.770	0.768
M	0.986	0.999	0.993
N	0.980	0.981	0.981
O	0.975	0.980	0.978
P	0.991	0.991	0.991
Q	0.890	0.928	0.908
R	0.979	0.978	0.978
S	0.992	0.989	0.991
T	0.975	0.988	0.981
U	0.961	0.944	0.952
V	0.945	0.958	0.951
W	0.997	0.990	0.993
X	0.990	0.991	0.991
Y	0.965	0.978	0.971
Z	0.995	0.998	0.996
blank	1.000	1.000	1.000

Hardest cases are I (F1=0.761) and L (F1=0.768), which are visually similar across handwriting styles. All other letters achieve F1 ≥ 0.90, and blank is perfect.

---

Usage

Alphanumeric model

from huggingface_hub import hf_hub_download
import onnxruntime as ort
import numpy as np
from PIL import Image

CHAR_CLASSES = list("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ") + ["blank"]

path = hf_hub_download(
    repo_id="hermitkk/alphabet-classifier",
    filename="outputs/exports/alphanumeric_model.onnx",
)
session = ort.InferenceSession(path)

# Preprocess a 96x96 grayscale crop
img = Image.open("character.png").convert("L").resize((96, 96))
x = (np.array(img, dtype=np.float32) / 255.0 - 0.5) / 0.5
x = x[np.newaxis, np.newaxis, :, :]  # (1, 1, 96, 96)

logits = session.run(None, {"input": x})[0]
pred = int(np.argmax(logits))
print(CHAR_CLASSES[pred])  # e.g. "A", "3", "blank"

Alphabet model

from huggingface_hub import hf_hub_download
import onnxruntime as ort
import numpy as np
from PIL import Image

ALPHA_CLASSES = list("ABCDEFGHIJKLMNOPQRSTUVWXYZ") + ["blank"]

path = hf_hub_download(
    repo_id="hermitkk/alphabet-classifier",
    filename="outputs/exports/alphabet_model.onnx",
)
session = ort.InferenceSession(path)

# Preprocess a 96x96 grayscale crop
img = Image.open("letter.png").convert("L").resize((96, 96))
x = (np.array(img, dtype=np.float32) / 255.0 - 0.5) / 0.5
x = x[np.newaxis, np.newaxis, :, :]  # (1, 1, 96, 96)

logits = session.run(None, {"input": x})[0]
pred = int(np.argmax(logits))
print(ALPHA_CLASSES[pred])  # e.g. "A", "blank"

Note: both models accept 96 × 96 single-channel float32 input, normalized to mean 0.5 / std 0.5. White pixels (blank paper) map to +1.0 and dark ink maps toward −1.0.

---

Reproduce Training

git clone https://huggingface.co/hermitkk/alphabet-classifier
cd alphabet-classifier
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Train alphanumeric model
python main.py train --config config/config.yaml

# Train alphabet model
python main.py train --config config/config_alphabet.yaml

---

License

MIT