Sign in to view source links and access this dataset
Description
108,403 image crops derived from the Jawildtext dataset for Japanese scene-text recognition. Each crop is a perspective-warped, rectified text region from a source image, horizontally aligned for model training. The dataset was created by user 'nagohachi' and was last updated on May 26, 2026.
Use Cases
Train Japanese scene-text recognition models based on the pre-processed, rectified image crops.
Evaluate OCR model performance on real-world Japanese text images based on the provided transcriptions.
Fine-tune multilingual OCR systems for Japanese text based on the scene-specific image data.
Benchmark text detection and recognition algorithms using the standardized crop format.
Strengths
Contains 108,403 individual text image samples.
Crops are pre-processed to be tight and horizontally aligned, which likely reduces preprocessing effort for model training.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
Derived from the 'llm-jp/jawildtext' dataset on Hugging Face.
Collection Method
For each text polygon in the source data, the source image was perspective-warped onto a rectified bounding rectangle.
Freshness
Last updated 2026-05-26 16:49:21; freshness should be verified.
License is unknown; terms of use must be verified before application.