AsterOCR: Vietnamese Neural Optical Character Recognition Engine

AsterOCR is a non-autoregressive neural Optical Character Recognition engine designed specifically for Vietnamese typography, administrative records (Citizen Identity Cards, passports, driving licenses), financial invoices, and smartphone document captures.

AsterOCR replaces legacy autoregressive Transformer models with an end-to-end convolutional-recurrent architecture accelerated via ONNX Runtime and PyTorch, delivering sub-60ms single-line inference on standard CPUs.

---

1. Architectural Blueprint and Parameter Profile

AsterOCR eliminates iterative autoregressive token decoding in favor of an end-to-end non-autoregressive sequence architecture.

graph TD
    A["Input Image Strip (32 x W x 3)"] --> B["HP-Stem: High-Pass Spatial Feature Filter"]
    B --> C["Backbone: Inverted Residual SE-ResNet (2.53M Params)"]
    C --> D["Map-to-Sequence (Feature Map -> 1D Temporal Slices)"]
    D --> E["Bidirectional LSTM (2 Layers, 256 Dim - 3.15M Params)"]
    E --> F["Multi-Head Self-Attention Fusion (4 Heads - 4.21M Params)"]
    F --> G["CTC Prediction Head (10.65M Total Params)"]
    G --> H["Neuro-Symbolic Linguistic Rescorer (Phonology & N-gram)"]
    H --> I["Output Vietnamese Text (Unicode NFC Standard)"]

Complete Parameter Profile (10,649,106 Total Parameters)

Neural Component	Module Architecture	Parameter Count	Ratio (%)	Operational Role
Input Stem	High-Pass Spatial Laplacian Filter + Conv3x3	Included in Backbone	—	Preserves micro 1-2px Vietnamese accents (ă, â, ơ, ư)
Vision Backbone	Inverted Residual MobileNetV3 + SE Channel Gating	2,525,249	23.7%	Spatial feature extraction and background noise suppression
Sequence Modeler	2-Layer Bidirectional LSTM (hidden_size=256)	3,153,920	29.6%	Bidirectional typographical stroke context dependencies
Language Fusion	4-Head Linear Multi-Head Self-Attention	4,206,592	39.5%	Long-range compound word coherence across sentence
Output Classifier	Linear Projection (hidden_dim -> 227 Classes)	116,451	1.1%	CTC probability distributions across character vocabulary
Auxiliary Heads	Multi-Task Tone and Diacritic Hat Classifiers	646,894	6.1%	Explicit multi-task supervision during representation learning
TOTAL	Full AsterNet Architecture	10,649,106	100.0%	40.76 MB ONNX Model (72% smaller than VietOCR 145MB)

---

2. Benchmark Accuracy Comparison vs VietOCR

Evaluated on 100 representative lines of the National Vietnamese Benchmark Suite (vi_00 / InkData):

Benchmark Metric	AsterNet v1.0	VietOCR (Transformer)	Delta / Gain
Exact Sentence Match (100% Match)	86.0%	76.0%	+10.0% accuracy gain
Character Error Rate (CER)	3.45%	3.85%	1.1x lower character error
Word Error Rate (WER)	9.01%	8.23%	Comparable word accuracy
Average Latency (Intel CPU)	60.25 ms	765.70 ms	12.8x faster inference
Throughput (Single CPU)	16.6 FPS	1.3 FPS	Suitable for live video feeds
Model Size	40.76 MB	~145 MB	72% lighter footprint

---

3. Real-World CPU Performance and Hardware Matrix

AsterOCR is engineered for lightweight execution without requiring dedicated GPU infrastructure:

Hardware Platform	Architecture	Threads	Latency / Line	Throughput (FPS)	Feasibility
Apple M3 / M4 Pro	ARM64 (NEON)	4	28.4 ms	35.2 FPS	Ultra Real-Time
Apple M1 / M2	ARM64 (NEON)	4	34.8 ms	28.7 FPS	Ultra Real-Time
Intel Core i7-13700H	x86_64 (AVX2/FMA)	1	60.2 ms	16.6 FPS	Standard Desktop / Server
AMD Ryzen 7 7840HS	x86_64 (AVX-512)	1	54.1 ms	18.5 FPS	Standard Desktop / Server
Cloud vCPU (AWS c6i.large)	x86_64 (AVX-512)	2	62.7 ms	15.9 FPS	Low-Cost Cloud Instance
Raspberry Pi 5 (4GB)	ARM Cortex-A76	4	118.0 ms	8.5 FPS	Edge / IoT Hardware
NVIDIA RTX 3060 / T4	CUDA FP16	Batch 16	5.8 ms	275.0 FPS	High-Concurrency API Server

---

4. Mathematical Formulations

4.1. High-Pass Spatial Feature Stem (HP-Stem)

Standard convolutional downsampling kernels act as low-pass filters, frequently obliterating subtle Vietnamese diacritics. HP-Stem introduces a learnable residual high-pass kernel $K_{hp}$:

$$F_{stem}(X) = \text{Conv}{3\times3}(X) + \gamma \cdot (X * K{hp})$$

where $\gamma$ is a dynamic scalar parameter amplifying edge gradients of high-frequency diacritic strokes before entering the pooling pyramid.

4.2. Inverted Residual with Squeeze-and-Excitation (SE-ResNet)

To suppress background paper discoloration, scanner artifacts, and watermark textures, AsterNet incorporates SE channel gating:

$$z_c = \frac{1}{H \times W} \sum_{i=1}^H \sum_{j=1}^W u_c(i, j)$$

$$s = \sigma\left(W_2 \cdot \text{ReLU}(W_1 z)\right)$$

$$\tilde{X}_c = s_c \cdot u_c$$

4.3. Contrast-Based Auto-Polarity Engine

Traditional CTC decoders fail on inverted text strips (white text on dark backgrounds, camera timestamps, dark mode). AsterOCR implements a mathematical contrast polarity check before tensor normalization:

$$\Delta_{bright} = P_{90}(X_{center}) - \text{Median}(X_{border})$$

$$\Delta_{dark} = \text{Median}(X_{border}) - P_{10}(X_{center})$$

$$\text{If } (\Delta_{bright} > \Delta_{dark} \text{ and } \Delta_{bright} > 20) \implies X \leftarrow (255 - X)$$

---

5. Quickstart (Python)

5.1. Installation

pip install asterocr

(No heavy PyTorch installation required by default. Runs using high-performance ONNX Runtime.)

5.2. Python API

from asterocr import AsterOCR

# Initialize engine (automatically fetches ONNX weights on first run)
ocr = AsterOCR()

# Recognize text from file path, PIL Image, bytes, or numpy array
text = ocr.predict("path/to/invoice.jpg")
print(text)
# Output: CỘNG HÒA XÃ HỘI CHỦ NGHĨA VIỆT NAM

5.3. Batch Recognition

results = ocr.predict_batch([
    "line_01.jpg",
    "line_02.jpg",
    "line_03.jpg"
])
for line in results:
    print(line)

5.4. Command Line Interface (CLI)

asterocr path/to/image.jpg

---

6. Multi-Platform Distribution Ecosystem

Platform	Package	Runtime Engine	Repository / Registry
Python	asterocr	onnxruntime / torch	pypi.org/project/asterocr
Node.js	@asterxlabs/asterocr	onnxruntime-node	npmjs.com/package/@asterxlabs/asterocr
Flutter	asterocr_flutter	onnxruntime_flutter	github.com/asterxlabs/asterOCR
Model Registry	asternet-vietnamese-ocr	ONNX / PyTorch	huggingface.co/asterxlabs/asternet-vietnamese-ocr

---

7. Model Artifacts and Verification

• asternet.onnx (40.76 MB): Production ONNX runtime model for cross-platform execution.
• asternet_best.pth (40.78 MB): Production PyTorch checkpoint trained on 1.56M Vietnamese samples.
• vocab.json: 227-character vocabulary mapping tokens to characters.
• config.json: Model architecture specifications and normalization hyperparameters.

---

8. License and Citation

AsterOCR is released under the MIT License. Free for commercial and non-commercial applications.

Developed by asterxlabs (2026).