Box the text
Drag a box around each text line or word. Four-point quadrilaterals are supported for skewed text: drag a corner after drawing.
Annotation
OCR Text Annotator
Made for training text detection and recognition models. Draw a box around each line or word, and either type the text or press Suggest to run Tesseract on the crop in your browser and correct the result. Exports include the PaddleOCR label file, ICDAR-style ground truth per image, and a recognition dataset of cropped images with a labels file.
Loading the tool…
How it works
Transcribe
Type into the field or press Suggest; Tesseract (English and Hindi bundled) reads the crop locally. Tab moves to the next box.
Export
PaddleOCR: Label.txt with JSON points and transcription per image. ICDAR: gt_<image>.txt with x1,y1,...,x4,y4,text. Recognition: crops/ folder plus rec_gt.txt.
Formats, limits and the numbers that matter
PaddleOCR layout
Label.txt lines are img/<name>\t[{"transcription": "...", "points": [[x,y],[x,y],[x,y],[x,y]]}, ...]. Detection training reads this file directly; recognition training reads rec_gt.txt with crop path and text.
Crops
Recognition crops are cut from the original image at full resolution, perspective-corrected for quadrilaterals, and saved as PNG.
Suggestions
Tesseract runs on the crop at 2x scale with a single-line page mode. It is a starting point, not ground truth: check every suggestion before export.
Languages
English and Hindi models are included. Other scripts can be typed by hand; suggestions for them are not available offline.
Common questions
Line or word boxes?
Match the model: PaddleOCR detection is trained on text lines; word-level boxes suit scene-text models like CRAFT. Pick one convention and keep it.
How is illegible text handled?
Mark the box as ignore; PaddleOCR writes it with transcription ### so the detector learns the region but recognition skips it.
Can I import an existing Label.txt?
Yes, drop it with the images; boxes and transcriptions are restored for editing.