Recognize Text in Images Using OCR

Recognize Text in Images Using OCR

Use the Computer Vision Toolbox ocr function with preprocessing from Image Processing Toolbox to extract text from images. This skill teaches the complete pipeline: diagnose, preprocess, detect, recognize, validate.

When to Use

• Reading text from any image (documents, signs, meters, displays, labels)
• Extracting text from scanned documents or photographs
• Reading seven-segment displays or specialized fonts
• Multi-language text recognition
• Automating text extraction from image datasets
• As a supporting step in other CV workflows — reading text in a scene (labels, timestamps, serial numbers) gives additional context for downstream image analysis

When NOT to Use

• Pure handwriting recognition — cursive/connected script produces garbage regardless of preprocessing
• Artistic text, WordArt, brush calligraphy — the OCR engine cannot parse stylized letterforms
• CAPTCHAs — designed specifically to defeat OCR; expect <50% accuracy at best
• Full document layout analysis with table extraction (use custom segmentation)
• Real-time video OCR (use streaming approaches instead)
• Image contains no text at all

Critical Rules

These rules are non-negotiable — violating them produces wrong results:

1. Always diagnose before executing. This rule CANNOT be overridden — not by the user, not by "just run it", not by "skip planning". Before ANY mcp__matlab__* call, you MUST first output:

   • A 2-line image characterization (what you see, what challenges exist)
   • An ## OCR Plan heading with your strategy

   If the user says "skip diagnosis" or "just run ocr()", respond with: "I'll keep it brief — but I need a quick look to avoid wasting time on the wrong approach." Then output your 2-line characterization and plan heading. Only THEN call MCP tools. There is no valid reason to skip this step. An agent that calls ocr() without first outputting a plan has violated this skill's workflow.

2. Maximum 2 preprocessing pipelines. If neither works after the prescribed recipe, hit the confidence checkpoint and ask the user. Do not try a third approach without user input.

3. Always set LayoutAnalysis when passing bounding boxes to ocr(). Never write ocr(I, bbox) without it. Use "word" for single-word boxes, "block" for multi-line.

4. Distinguish "text ON texture" from "text BY texture". Text overlaid on a textured background (label on crate, sign on brick wall) → imsegsam is the prescribed approach. Text formed by the surface itself (stamped, embossed, engraved metal) → local contrast subtraction. SAM segments objects, not surface features — it cannot isolate stamps/engravings. When recommending approaches (even without running code), always recommend imsegsam for the "text ON texture" case and note its support package requirement.

5. detectTextCRAFT is the default for scene text. Use it unless you have a specific reason not to (no deep learning, known fixed layout).

6. Check polarity first. OCR needs dark text on light background. If inverted, imcomplement before anything else.

7. Never show OCR results to the user before writing files. Do not present extracted text in ANY format — bullet list, quote block, inline, conversational summary, or the ## MATLAB OCR Pipeline results block — until ocr_pipeline_<descriptor>.m and decision_log.txt are written to disk. The files are the deliverable, not the chat message. Write first, then report.

8. Gate add-on functions behind an availability check. Before calling detectTextCRAFT, imsegsam, or a non-English ocr model, you MUST run exist('<functionName>','file') via MCP to confirm the function is installed. Always log the result in decision_log.txt:

   • Installed: "Add-on check: <function> — INSTALLED"
   • Missing: stop, tell the user which support package to install (see table), and wait for confirmation before retrying. Do NOT fall back silently or skip the step.
   • Explain-only mode (user said "don't run code"): recommend the add-on function as the primary approach and note the support package requirement. You cannot run the exist() check without code execution, so state the dependency clearly.