An OCR result can look convincing while changing one digit in an identifier or one character in a name. A confidence score helps prioritize review, but it is not a calibrated probability that a business fact is correct. The consumer needs both the recognized text and enough page evidence to inspect the uncertain part.
Image OCR Text Extractor recognizes printed English and Russian from authorized PNG, JPEG and scanned PDF inputs. It uses Tesseract with installed language models. It does not translate the text, infer a document's meaning or promise handwriting recognition.
Begin with a controlled page
The installed prefill is an owned English image fixture. Request one page and inspect its text, word boxes and confidence fields before importing a mixed archive. The schema example below sets the page and pixel bounds explicitly.
{
"fileUrls": [
"https://api.apify.com/v2/key-value-stores/b0cgFsJRqVv3XOZyG/records/ocr-en.png"
],
"languages": [
"eng"
],
"maxPagesPerFile": 1,
"maxPixelsPerPage": 5000000
}
One through three source URLs are supported. languages selects the installed language models, with up to two selected languages. Choose models that match the source rather than assuming a language selection will translate unfamiliar text into English.
Keep boxes and page identity with the words
The output includes word bounding boxes, per-word confidence, mean confidence, actual image dimensions and engine version. Preserve those fields with the source page identity. A reviewer can then locate a questionable identifier on the image instead of searching an unstructured block of text.
If your consumer extracts fields from the recognized text, store the field proposal separately from the raw OCR output. Correcting a proposed invoice identifier should not overwrite the original recognition evidence. The two records explain different stages of the workflow.
Set a review policy around the task
A low-confidence word can be harmless in a paragraph and critical in a reference number. Validate important identifiers against their expected formats and route uncertain cases for inspection. A high mean page confidence does not guarantee that the single most important token is correct.
Review thresholds should be based on representative pages and the consequences of an error. The Actor's confidence values are engine scores, not a universal acceptance policy. Keep the original image available for authorized reviewers when the consumer cannot resolve a doubtful value mechanically.
Respect page-level completion
The input permits at most five pages per file and at most five million pixels per page. A PDF is rasterized under the documented 150 DPI policy before recognition. Text from a complete nonempty page can remain useful even if a later page fails or a selected page cap leaves a remainder.
Preserve the separate partial or error notice describing that remainder. Counting every dataset row as a complete recognized page would incorrectly count operational notices. Likewise, one good page does not prove that the entire source document was processed.
Keep blank pages and failures distinct
A blank page can legitimately produce no text. An oversized image, timeout or failed download is a different outcome. Keep these states in the ingestion log so an operator can decide whether to retry, resize an authorized source or request a better scan.
OCR is most useful when its uncertainty stays attached to the output. The goal of this workflow is a traceable recognition and review process, with page evidence and explicit missing scope, rather than a text field that appears certain because its provenance was discarded.
Cost and verification scope
The current FREE-tier result price is $0.008 for the documented unit: One OCR-processed image or PDF page with recognized text. The listing also has a $0.005 start event, charged per GB of Actor memory with a minimum of one event. A free diagnostic row does not refund that start event. Discounted platform tiers have their own rates; inspect your applicable tier before estimating a batch. These are listing prices read on 21 October 2026, not a receipt or a claim about customer spending.
The contract for this draft was checked against installed build 0.1.2 (olzihcrtVZdDG5Ngk). The input above is an illustrative schema example, not a newly executed run. Current source responses and useful output still need inspection when you choose to run it. No new cloud build or run was made to prepare this article.
Open the Actor and inspect its input and pricing.
AI-assisted draft; publication requires independent technical review.
Top comments (0)