Image to Text - OCR, Word Coordinates & 图片文字识别
Pricing
$20.00 / 1,000 nonempty images
Image to Text - OCR, Word Coordinates & 图片文字识别
Read printed text from public image URLs with OCR. One row per nonempty image: text, words, confidence, bounding boxes, language and TXT download URL. 图片文字识别。 English, simplified Chinese or both. Free empty images and ID/hash exclusions; no handwriting guarantee.
Pricing
$20.00 / 1,000 nonempty images
Rating
0.0
(0)
Developer
Tedj MEABIOU
Maintained by CommunityActor stats
0
Bookmarked
2
Total users
1
Monthly active users
7 days ago
Last modified
Categories
Share
Image to Text provides ocr and downloads for automated workflows, with JSON export, repeat monitoring and bounded pay-per-result delivery. Extract bilingual text from screenshots and document images for agents that need searchable text plus evidence coordinates.
Image to Text: scope and source behavior
Uses the open-source Tesseract OCR engine. Pass eng for English, chi_sim for simplified Chinese, or eng+chi_sim for mixed text. Default English does not imply automatic language detection. Output includes text assembled by detected lines and word bounding boxes in pixels. Confidence is the engine score, not a calibrated probability. Printed, high-contrast upright text works best; handwriting, low resolution, complex typography and skew can produce errors. Each image is limited to 20 million pixels and 12 MB compressed download. The first frame of a multi-frame image is processed. At most 20,000 words are returned; truncated flags longer output. There is no PDF input or structured invoice interpretation.
This is an independent Kestrel utility. It is not an official product of the source, registry, browser vendor or document library. It accepts the documented inputs and returns observed or extracted data with source references. It does not create source accounts, buy data subscriptions, send messages, alter websites or submit forms. The exact supported operation matters more than a generic promise to handle every possible agent request.
Image to Text for AI agents: which input answers which request?
Use this Actor when the requested output matches the scope above. Read the input schema before constructing a call. Required target values are explicit lists, while optional settings have documented defaults. A run with omitted target fields uses the demonstration values; production callers should always send their intended targets. A source account or source API key is not required for the supported operation.
| Request | Input approach |
|---|---|
| Try the documented example | Use the small JSON sample below. |
| Process a batch | Put up to 100 targets in imageUrls. |
| Return at most three paid results | Set maxResults to 3. |
| Skip previously delivered identities | Copy previous id values into excludeIds. |
| Skip previously seen snapshots | Copy previous content_hash values into excludeContentHashes. |
| Limit spending | Set maxTotalChargeUsd in the Apify run options, outside input JSON. |
The example is a small test, not a promise of a fixed row count. A document may contain empty pages, a website may display no email address, or an image may have no recognizable text. Valid empty extractions are distinguished from upstream errors. Keep the free summary alongside the dataset so your agent can explain what actually happened.
OCR inputs and limits
| Input | Behavior |
|---|---|
imageUrls | Public image URLs, e.g. the prefilled Tesseract sample. Up to 100; 12 MB and 20 million pixels each. One billed row per nonempty OCR image. |
language | "eng" (default), "chi_sim" for simplified Chinese or "eng+chi_sim" for mixed text. Not automatic detection. Language selection has no extra fee. |
maxResults | Maximum paid rows across the run, 1–1000, default 20. Example: 3 for a sample. Exclusions and empty results apply before the cap. |
excludeIds | Prior output id values to skip before billing, up to 5,000. Empty [] disables. Suppresses that identity even if content changed. |
excludeContentHashes | Prior output content_hash values (64 lowercase hex characters), up to 5,000. Empty [] disables. Matching snapshots are skipped before billing. |
All fields reject incompatible types. Boolean values must be JSON booleans, not strings. Lists contain strings rather than nested request objects. Unknown property names fail validation to catch misspelled agent inputs early. Duplicate target strings are removed within a run. Domain names are normalized before deduplication. Exclusion lists are optional and bounded so a single run cannot grow without limit.
The overall result limit applies after exclusions. A qualifying result removed by an exclusion does not consume that result allowance. Source page limits, download limits and processing limits remain independent: excluding an old result does not refund the work needed to discover it. A request for twenty rows can therefore finish with fewer than twenty rows without indicating a bug.
Downloads output and download contract
One paid row represents one nonempty image. The principal observed row fields are shown below. Optional source fields can be missing or null; consumers should not invent replacements. The output remains structured JSON even when a downloadable file is also produced.
| Field | Meaning |
|---|---|
text | text |
words | words |
truncated | truncated |
language | language |
source_url | source url |
id | Stable identity within this Actor and target scope. |
content_hash | SHA-256 of the normalized result for comparison. |
fetched_at | UTC extraction or observation time. |
Document conversion, OCR, website Markdown and screenshots include download_url, artifact_key, artifact_bytes and artifact_sha256 when a file is delivered. Those fields identify a file in the run's Apify key-value store. Do not remove the signature query parameter from a generated download URL. The generated link includes a record-specific signature for downloading without an Apify token. Files remain available only while the store and record are retained. Export important artifacts into your own storage rather than treating a run as a permanent archive.
Email, DNS, registration and weather rows are dataset records and do not create a separate file download. Use Apify dataset export to obtain JSON or CSV where useful. Nested values such as tables, words, entities and forecasts are better preserved as JSON than flattened CSV. A download does not contain more coverage than the associated row; a PDF page workbook, for example, does not silently include unprocessed document pages.
What does a successful run mean?
A successful run means the requested bounded work finished without a reported target error. It does not guarantee complete website coverage, perfect OCR, a reachable mailbox, future weather accuracy or a registry field that the source did not provide. Each product's limitations are specified in the scope section. Callers should verify a representative output before placing it into a consequential downstream workflow.
The free SUMMARY record contains status, rows, excluded, targets_attempted, targets_requested, stop_reason, errors, artifacts and coverage. It also reports source_http_requests (including redirects and failed connection attempts), source_downloaded_bytes (decoded response bytes) and dns_cache_hits for workload inspection. These counters exclude Apify storage traffic and system DNS resolution. The row count is the number delivered into the paid dataset. The artifact list contains file keys only for successfully delivered file-bearing rows. A budget or result limit can stop the run before every requested target is attempted.
Status no_matches means no qualifying row was produced without a target failure. Status error means a failure with no valid delivered row, and partial_error means valid rows were delivered before or alongside a failure. Error states fail the Actor run so automation does not mistake an unavailable source for an empty result. Invalid inputs are reported as invalid_input before any target access.
How much does it cost?
The event price is $0.02 per delivered nonempty image, equivalent to $20 per 1,000 rows. Ten rows cost $0.2; one hundred cost $2. There is no Actor start fee. Compute and network access are included in the event price. Nested data and files described for a row are included in that row's price.
Within-run duplicates, excluded identities, excluded matching snapshots and empty extractions are free. Failures without delivered data produce no result charge. An invalid-syntax email is a valid diagnostic result and is billed as an email check; a DNS NXDOMAIN response and a registry not-found response are likewise useful observations, not failed requests. The Actor never charges a separate event for every nested word, cell, DNS record or forecast step.
Set maxTotalChargeUsd in Apify run options for a monetary ceiling. It is not an Actor input field. The charging manager is checked before starting a target and before delivering a result; delivery stops when no complete result event fits the remaining budget. Previously delivered rows remain available and retain their charges if a later target fails. Customer event prices and developer infrastructure costs are different figures; a test run's compute cost is not a customer revenue measurement.
How do Python, JavaScript and curl integrations work?
The Python example uses apify-client 3.x and its typed Run object. Use an environment variable for the Apify token and the asynchronous run API for reliable orchestration. Start the run, wait for its terminal status, then read both the dataset and the free summary. Keep the run identifier in downstream records so an extraction can be traced to its inputs, build and status. Never paste credentials into target URLs.
How should repeat monitoring work?
Persist returned identities and hashes in your own database. On subsequent runs, supply the relevant exclusion list. An identity exclusion suppresses that target or page even if its content changes. A hash exclusion suppresses only a previously seen normalized state. Choose hashes for change monitoring and identities for one-time collection. The Actor does not maintain a hidden shared contact history across customers.
The hash is for comparison, not proof of authenticity. It excludes the observation timestamp and storage download URL. Screenshot image bytes contribute to their hash, so dynamic banners, animations or dates can make a new screenshot differ even when the site's main content is unchanged. DNS TTLs and evolving forecasts also make snapshots change. Use field-specific downstream comparisons when your business definition of change is narrower.
Batch related targets and choose a schedule that matches the source update cadence. A repeated run cannot force a registry, weather model or website to publish new information. Avoid retry loops on permanent input errors. A transient upstream failure can justify a later bounded retry, with exclusions for rows already delivered. There is no built-in notification, email sending or schedule creation; those actions belong to the calling workflow.
How do n8n, Make and MCP agents use this?
Use an Apify connector or an HTTP request module to start the run and wait for completion. In n8n and Make, branch on the run status before iterating rows. Download file artifacts only when a row actually contains an artifact key. Store the summary on both successful and failed branches. This makes partial results usable without hiding unobserved targets.
An MCP agent should inspect the Actor details, choose literal input fields and begin with a small sample. It should quote observed values and source URLs, preserve units and page numbers, and describe source limitations when they affect the answer. A structured result is easier to use correctly than an undocumented text blob, but it still requires interpretation. Do not treat a confidence score as certainty or a missing field as a negative fact.
FAQ: common operational questions
Does this accept authenticated sites or private-network URLs?
No. Only public HTTP(S) targets on standard ports are supported for website and document inputs. Embedded credentials, loopback, private, link-local and reserved addresses are rejected. DNS resolution is checked before connections, and redirects are revalidated. The screenshot browser routes public GET resources through the same bounded fetch layer. This protects the execution environment and keeps the supported input contract predictable.
Does it translate or invent missing content?
No. Chinese discovery text helps users find the utility, while output values retain their source meaning. OCR language selection controls recognition, not translation. A missing field stays missing or null. Document extraction does not infer a table that was not detected, and website conversion does not summarize text with an LLM. Downstream agents can translate or analyze the output separately while retaining the original evidence.
What happens with a large or malformed file?
Downloads are capped, and parsing takes place in a separate process with a wall-time limit. An oversized, encrypted or malformed document can fail explicitly. This is preferable to returning a silently incomplete conversion. The supported page and image limits are documented in the input descriptions and scope. Split large batches into smaller runs so one problematic target is easy to isolate and retry.
Are results deterministic?
The same stable source and settings usually produce equivalent normalized text or structured data, but upstream pages and endpoints change. Browser screenshots also depend on timing, responsive layout, fonts and dynamic content. Workbook binary metadata can differ even when extracted table cells match; the comparison hash uses normalized cells rather than requiring byte-identical workbooks. Preserve both hashes and source references when building an audit trail.
Can I claim completeness from an empty dataset?
No. Read the run status and summary first. A genuinely empty extraction, a scan cap and an upstream failure are different outcomes. An email-free starting page does not establish that a company has no email address. A PDF page without extractable text may be a scan. A missing registry object does not establish that a domain is available to buy. The utility reports the observed operation, not a broader unsupported conclusion.
What legal and source-use considerations apply?
Use public data and documents for purposes you are authorized to pursue, and retain applicable source notices. Public availability does not remove copyright or database restrictions. Source names identify the tool or service used and do not imply endorsement. MET Norway data requires attribution and its license reference, which the forecast row includes. Registry notices are preserved. The utility does not grant rights over supplied documents or collected website content.
中文说明：输入、结果与费用
本工具支持图片文字识别、中文 OCR、图片转文字。输入参数名保持英文，便于 AI 智能体和 MCP 客户端正确调用。结果包含来源、采集时间、稳定标识和内容哈希。请先使用少量目标验证结果，再扩大批量任务。
每条实际交付的结果计费；重复、排除和空提取结果不收费。请在 Apify 运行选项设置 maxTotalChargeUsd，不要放进输入 JSON。maxResults 限制整个任务的付费结果数量。文件转换工具会提供下载链接和文件键，访问权限由 Apify 存储设置决定。
本工具不会编造缺失字段，也不保证源站完整覆盖。邮箱 DNS 检查不等于邮箱可投递验证；数字 PDF 提取不等于扫描件 OCR；天气预报不等于历史实测数据。失败任务与真正的空结果明确区分，请同时检查数据集及免费的 SUMMARY。
Source documentation: https://tesseract-ocr.github.io/tessdoc/
Last verified working: 2026-10-17
