---

Paper | Model Download | Dataset Download | Inference Space |

A Squeeze-and-Excitation Transformer Network for Khmer Optical Character Recognition

Character Error Rate (CER %) on KHOB, Legal Documents, and Printed Word Benchmark

Introduction

This repository contains the implementation, datasets, and evaluation results for the Squeeze-and-Excitation Transformer Network, a high-performance Khmer Text Recognition model that utilizes a hybrid architecture combining Squeeze-and-Excitation blocks for feature extraction and BiLSTM smoothing for context smoothing, specifically designed to handle the complexity and length of Khmer script.

Overview

Khmer script presents unique challenges for OCR due to its large character set, complex sub-consonant stacking, and variable text line lengths. This project proposes an enhanced pipeline that:

1. Chunks long text lines into manageable overlapping segments.
2. Extracts Features using a Squeeze-and-Excitation Network (SE-VGG) that preserves horizontal spatial information.
3. Encodes local spatial features using a Transformer Encoder.
4. Merges the encoded chunks into a unified sequence.
5. Smooths Context using a BiLSTM layer to resolve boundary discontinuities between chunks.
6. Decodes the final sequence using a Transformer Decoder.

Datasets

The model was trained entirely on synthetic data and evaluated on real-world datasets.

Training Data (Synthetic)

We generated 200,000 synthetic images to ensure robustness against font variations and background noise.

Dataset Type	Count	Generator / Source	Augmentations
Document Text	100,000	Pillow + Khmer Corpus	Erosion, noise, thinning/thickening, perspective distortion.
Scene Text	100,000	SynthTIGER + Stanford BG	Rotation, blur, noise, realistic backgrounds.

Evaluation Data (Real-World + Synthetic)

Dataset	Type	Size	Description
KHOB	Real	325	Standard benchmark, clean backgrounds but compression artifacts.
Legal Documents	Real	227	High variation in degradation, illumination, and distortion.
Printed Words	Synthetic	1,000	Short, isolated words in 10 different fonts.

Methodology & Architecture

1. Preprocessing: Chunking & Merging

To handle variable-length text lines without aggressive resizing, we employ a "Chunk-and-Merge" strategy:

• Resize: Input images are resized to a fixed height of 48 pixels while maintaining aspect ratio.
• Chunking: The image is split into overlapping chunks (Size: 48x100 px, Overlap: 16 px).
• Independent Encoding: Each chunk is processed independently by the Squeeze-and-Excitation Network and Transformer Encoder to allow for parallel batch processing.

2. Model Architecture: Squeeze-and-Excitation Transformer Network

Our proposed architecture integrates sequence-aware attention and recurrent smoothing to overcome the limitations of standard chunk-based OCR. The model consists of six key modules:

1. Squeeze-and-Excitation Network (SE-VGG):

   • A modified VGG backbone with 1D Squeeze-and-Excitation blocks after convolutional layer 3, 4, and 5.

   • Unlike standard SE, these blocks use vertical pooling to refine feature channels while strictly preserving the horizontal width (sequence information).

     

2. Patch Module:

   • Projects spatial features into a condensed 384-dimensional embedding space.
   • Adds local positional encodings to preserve spatial order within chunks.

3. Transformer Encoder:

   • Captures contextual relationships among visual tokens within each independent chunk.

4. Merging Module:

   • Concatenates the encoded features from all chunks into a single unified sequence.
   • Adds Global Positional Embeddings to define the absolute position of tokens across the entire text line.

5. BiLSTM Context Smoother:

   • A Bidirectional LSTM layer that processes the merged sequence.

   • Purpose: Bridges the "context gap" between independent chunks by smoothing boundary discontinuities, ensuring a seamless flow of information across the text line.

     

6. Transformer Decoder:

   • Generates the final Khmer character sequence using the globally smoothed context.

---

Training Configuration

• Epochs: 100
• Optimizer: Adam
• Loss Function: Cross-Entropy Loss
• Learning Rate Schedule: Staged Cyclic
  • Epoch 0-15: Fixed 1e-4 (Rapid convergence)
  • Epoch 16-30: Cyclic 1e-4 to 1e-5 (Stability)
  • Epoch 31-100: Cyclic 1e-5 to 1e-6 (Fine-tuning)
• Sampling: 50,000 images randomly sampled/augmented per epoch.

---

Quantitative Analysis

We benchmarked our proposed model against VGG-Transformer, ResNet-Transformer, and Tesseract-OCR.

Character Error Rate (CER %) - Lower is better

TABLE 1: Character Error Rate (CER in %) results on the KHOB, Legal Documents, and Printed Word

Model	KHOB	Legal Documents	Printed Word
Tesseract-OCR	6.24	24.30	8.02
VGG-Transformer	2.27	10.27	3.61
ResNet-Transformer	2.98	11.57	2.80
Proposed Model	$\textcolor{yellow}{1.87}$	$\textcolor{yellow}{9.13}$	$\textcolor{yellow}{2.46}$

---

Qualitative Analysis

TABLE 2: Failure cases on KHOB, Legal Document, and Printed Word dataset

TABLE 3: Example of proposed, and baseline model compared with the ground truth. Errors in the predictions are highlighted in red

Key Findings:

• The Proposed Model achieves the highest accuracy on long, continuous text lines (KHOB), demonstrating that the BiLSTM Context Smoother effectively resolves the chunk boundary discontinuities that limit standard Transformer baselines.
• On degraded and complex legal documents, the proposed model demonstrates superior robustness, significantly outperforming all baselines. This attributes to the Squeeze-and-Excitation blocks, which filter background noise while preserving character-specific features.
• The Proposed Model still retains a slight advantage on short, isolated words even where global context is less critical, outperforming both ResNet and VGG Transformer baseline.

Setup

Create virtual environment

# Windows
python -m venv myenv
.\myenv\Scripts\activate

# Mac/Linux
python3 -m venv myenv
source myenv/bin/activate

Installation

pip install -v git+https://github.com/netra-ai-lab/Khmer-OCR-CNN-Transformer.git@master

---

Inference Usage

This pipeline performs end-to-end OCR by first detecting text lines using Surya and then recognizing characters using our custom CNN-Transformer model.

Local-Inference

1. Command Line Interface (CLI)

netra_ocr --image path/to/your/image.jpg

You can also adjust detection engine, adjust beam, and enable debugging

netra_ocr --image document.png --engine surya --padding 10 --beam 1 --batch_size 16 --debug

CLI Arguments

Argument	Type	Default	Description
--image	str	Required	Path to the input image file.
--engine	str	surya	Choice of detection engine: surya (optimized for textlines) or custom (trained layout model).
--output	str	ocr_result.txt	Path to save the final recognized text.
--padding	int	6	Pixels of white-space padding added around each text line. Higher padding can help recognition accuracy.
--beam	int	1	Beam width for recognition. 1 is Greedy Search (fastest). Higher values increase accuracy but decrease speed.
--batch_size	int	8	Number of text lines to process in parallel on the GPU/CPU.
--debug	flag	False	If set, saves every cropped line image and its corresponding text into a debug_<filename>_<engine> folder for verification.

Huggingface-Inference

1. Setup

pip install torch torchvision transformers pillow huggingface_hub

# Setup the inference script
wget https://huggingface.co/Darayut/khmer-text-recognition/resolve/main/configuration_khmerocr.py

wget https://huggingface.co/Darayut/khmer-text-recognition/resolve/main/inference.py

2. Run via CLI

python inference.py --image "path/to/image.png" --method beam --beam_width 3

3. Run via Python

from inference import KhmerOCR

# Load Model (Downloads automatically)
ocr = KhmerOCR()

# Predict
text = ocr.predict("test_image.jpg", method="beam", beam_width=3)
print(text)

---

References

1. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
   Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al.
   ICLR 2021.
   arXiv:2010.11929

2. TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
   Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
   AAAI 2023.
   arXiv:2109.10282

3. Toward a Low-Resource Non-Latin-Complete Baseline: An Exploration of Khmer Optical Character Recognition
   R. Buoy, M. Iwamura, S. Srun and K. Kise.
   IEEE Access, vol. 11, pp. 128044-128060, 2023.
   DOI: 10.1109/ACCESS.2023.3332361

4. Balraj98. (2018). Stanford background dataset [Data set]. Kaggle. https://www.kaggle.com/datasets/balraj98/stanford-background-dataset

5. EKYC Solutions. (n.d.). Khmer OCR benchmark dataset (KHOB) [Data set]. GitHub. https://github.com/EKYCSolutions/khmer-ocr-benchmark-dataset

6. Em, H., Valy, D., Gosselin, B., & Kong, P. (2024). Khmer text recognition dataset [Data set]. Kaggle. https://www.kaggle.com/datasets/emhengly/khmer-text-recognition-dataset

7. Squeeze-and-Excitation Networks
   Jie Hu, Li Shen, and Gang Sun.
   CVPR 2018.
   arXiv:1709.01507

8. Bidirectional Recurrent Neural Networks
   Mike Schuster and Kuldip K. Paliwal.
   IEEE Transactions on Signal Processing, 1997.
   DOI: 10.1109/78.650093