Optical Character Recognition (OCR)

Optical Character Recognition (OCR) is a technology that converts different types of documents—such as scanned paper documents, PDF files, or images captured by a digital camera—into editable and searchable data. The core problem OCR solves is that images are pixel-based and lack semantic text information. Without OCR, the text in an image is just a collection of colored pixels; a computer cannot search, copy, or edit it. OCR bridges the gap between visual content and machine-readable text.

In the context of Azure AI, OCR is provided by the Computer Vision service, specifically through the Read API. The Azure AI Document Intelligence service (formerly Form Recognizer) also includes OCR capabilities, but its primary focus is on extracting key-value pairs and tables from forms. For the AI-900 exam, you need to understand the Read API as the primary OCR offering.

OCR in Azure AI follows a multi-step pipeline:

- Read API Endpoint: https://<your-resource-name>.cognitiveservices.azure.com/vision/v3.2/read/analyze (version 3.2 is commonly tested). - Operation: The Read API is asynchronous. You first call the POST endpoint to submit the image, which returns an Operation-Location header with a URL to poll for results. - Polling: You must poll the GET endpoint at the Operation-Location URL until the status is 'succeeded' or 'failed'. The recommended polling interval is 1 second, but not less than 0.5 second to avoid throttling. - Output Fields: The JSON response includes: - status: 'succeeded', 'running', 'failed' - analyzeResult: contains readResults array, each with: - page: page number (1-indexed) - angle: skew correction angle in degrees - width, height: dimensions of the image - unit: 'pixel' or 'inch' - lines: array of text lines, each with: - boundingBox: array of 8 numbers (x1,y1,x2,y2,x3,y3,x4,y4) - text: the recognized text - words: array of words, each with: - boundingBox: similar - text: the word - confidence: confidence score between 0 and 1 - Supported Languages: The Read API supports over 160 languages, including English, Spanish, French, German, Chinese, Japanese, Arabic, etc. For printed text, it supports many languages; for handwritten text, it supports English and a few others (like Spanish, French, German, Italian, Portuguese, Dutch). Check the latest documentation for the exact list. - Image Requirements:

To use the Read API, you need an Azure AI Services resource (formerly Cognitive Services) with the Computer Vision endpoint and key. You can call the API via REST or using SDKs (Python, C#, Java, etc.).

Example REST call:

POST https://<your-resource-name>.cognitiveservices.azure.com/vision/v3.2/read/analyze
Headers:
  Ocp-Apim-Subscription-Key: <your-key>
  Content-Type: application/json
Body:
{
  "url": "https://example.com/document.jpg"
}

Response header:

Operation-Location: https://<your-resource-name>.cognitiveservices.azure.com/vision/v3.2/read/analyzeResults/<operation-id>

Then poll:

GET https://<your-resource-name>.cognitiveservices.azure.com/vision/v3.2/read/analyzeResults/<operation-id>
Headers:
  Ocp-Apim-Subscription-Key: <your-key>

A retail company receives thousands of paper invoices daily from suppliers. They use Azure AI Document Intelligence (which includes OCR) to digitize invoices. The OCR step extracts all text from scanned invoices, then Form Recognizer extracts key fields like invoice number, date, total amount, and line items. The company processes about 10,000 invoices per day. They use the paid tier with a throughput of 100 TPS. They have a custom model trained on their invoice format. Common issues: poor scan quality (low resolution, skewed pages) leads to OCR errors. They preprocess images to ensure at least 300 DPI and deskew before sending. They also set up a fallback: if confidence scores are below 0.6 for critical fields, the invoice is sent for manual review. Misconfiguration: initially they used the Read API alone, but it did not extract structured data, so they had to switch to Form Recognizer.

A law firm receives physical mail that needs to be digitized and indexed. They use a scanner that automatically sends images to Azure Computer Vision OCR. The OCR extracts text from envelopes and letters. The text is then used to classify the mail (e.g., client, court, opposing counsel) using Azure Cognitive Search. They process about 500 documents per day. They use the free tier initially but upgrade to paid as volume grows. A common problem: handwritten addresses on envelopes have low accuracy (around 80%). They mitigate by using a custom handwriting recognition model trained on their specific handwriting styles (e.g., from known senders). They also integrate with Power Automate to trigger workflows based on recognized text. Misconfiguration: they initially used synchronous OCR calls, which caused timeouts for large documents. They switched to the asynchronous Read API and implemented proper polling.

A library digitizes historical books and newspapers. They use OCR to make the text searchable. The documents are old with faded ink and unusual fonts. They use the Read API with the 'language' parameter set to the document's language. They process about 100 pages per hour. To improve accuracy, they preprocess images with contrast enhancement and binarization. They also use the 'pages' parameter to process specific pages in a multi-page TIFF. A common issue: the OCR confuses similar characters like 's' and 'f' in old fonts. They use post-processing with a custom dictionary to correct known errors. They also store the raw OCR output for archival purposes, even if not perfect. Misconfiguration: they tried to process very large TIFF files (over 100 MB) that exceeded the file size limit. They now split files into individual pages before sending.