khmer-text-recognition

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.

Evaluation Data (Real-World + Synthetic)

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 CNN 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

Qualitative Analysis

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

TABLE 3: Example of our proposed model against all baseline 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,