PP-OCRv6 (ppocr-v6)

ppocr-v6 (PP-OCRv6 small) recognises text in screenshots and cropped regions. Each line is returned with its text, a confidence, a bounding box and a quadrilateral, and by default with word-level boxes as well. One call handles one image, whatever the number of lines.

Model ID	ppocr-v6 (default model of the endpoint)
Model	PP-OCRv6 small (PaddlePaddle)
Weights	Official PP-OCRv6 small detection and recognition models (PaddleOCR)
Use cases	Text recognition (OCR)
Endpoints	POST /v1/ocr
Pool	Public: available to every registered account, 100 calls per UTC day included
Billing units	image. Units and prices: Billing
Licence	Apache-2.0

How it works

A call runs two stages on the image: text detection, then line recognition. Billing follows the image, not these stages.

Detection. A detection model finds the regions of the image that contain text. Each region is a four-corner shape, so tilted lines are located as well as horizontal ones. The result is returned as the quad of the line and as an axis-aligned bbox.

Line recognition. A recognition model reads each detected region as one line and returns its text with a confidence. Lines below options.minConfidence (default 0.5) are dropped before the response is built.

Word boxes. With options.words (default true), the recogniser also locates the pieces of each line, and every line carries word boxes. The model gives no separate score per word, so each word carries the confidence of its line. With options.words set to false, the word boxes are left out and the response is smaller.

There is no layout analysis: lines come back as they were recognised, and grouping them into paragraphs or columns is left to the caller. For layout, see paddleocr-vl. Both stages run once per call and the work does not depend on a cache, so a call is billed as one image, however many lines or words are found.

In each use case

Text recognition (OCR)

ppocr-v6 is the default model of POST /v1/ocr, and the low-cost choice for screenshots, UI captures, receipts, code windows and cropped regions, and for any case where the text is to be copied out as it is.

• One image per call. Every text line found is returned with text, confidence, bbox and quad; the quad fits tilted text more closely than the box does.
• Word-level boxes are on by default and locate single tokens inside a line, for example to highlight a search hit or to place a click target. Whitespace can appear as a word of its own.
• options.minConfidence (0 to 1, default 0.5) drops weak lines.
• No layout, reading order, tables or formulas are returned.
• Coordinates are pixels of the oriented original: see Limits and notes.

Use case guide and example call: Text recognition (OCR).

Document parsing

Not offered by ppocr-v6: it returns plain lines, with no blocks, reading order, tables or formulas. Pages that need structure go to paddleocr-vl.

Limits and notes

• One image per call. Image limits: see Images.
• Stack resizes the image for the model before recognition. All coordinates are pixels of the oriented original, whose size is returned as image.width and image.height.
• Page orientation correction and unwarping are switched off: the image is read as it is sent.
• Latency: the recorded call (the example call of the use case page) had timing.processingMs of 753 ms. The figure depends on the image, on queueing and on node load, and it is not a guarantee.