Synthetic Document Images for English Word Recognition

MJSynth is a synthetic dataset for text recognition containing 90,000 English words on document images. It was created by the Visual Geometry Group at the University of Oxford and uploaded to HuggingFace by priyank-m in July 2023. The dataset includes predefined training, validation, and test splits.

Use Cases

• Train a text recognition model to predict the text label from synthetic document images.
• Benchmark model performance on the provided validation and test splits of image-text pairs.
• Fine-tune pre-trained vision-language models on the 90,000 word-image examples for improved OCR.
• Use the synthetic word images for data augmentation in document analysis pipelines.

Strengths

• Contains 90,000 distinct English words for model training.
• Provides structured training, validation, and test splits for evaluation.

Limitations

• Data is synthetically generated, which may not fully represent real-world document noise and variability.
• Limited to English words, reducing applicability for multilingual text recognition.

Provenance

Source

Visual Geometry Group, University of Oxford (https://www.robots.ox.ac.uk/~vgg/data/text/)

Collection Method

Synthetically generated document images.

Freshness

Dataset uploaded July 2023; original source creation date is unknown.