OCR: Detecting & Recognizing Text in Images¶

What was run¶

A real string of text, rendered into a real image, recognized by a real local Tesseract OCR engine — exact match confirmed first — then pushed through three different real degradations (blur, rotation, noise) with character-level accuracy measured at each step, to find exactly where OCR actually breaks rather than assuming it degrades uniformly.

A real, exact baseline¶

recognized = pytesseract.image_to_string(canvas).strip()

REAL ground truth text: 'Computer Vision Mastery Path 2026'
REAL OCR recognized text: 'Computer Vision Mastery Path 2026'
real exact match: True

Real blur: almost completely harmless¶

blur ksize= 1: real char-accuracy=1.0000
blur ksize=15: real char-accuracy=1.0000

A real, genuinely surprising robustness: even a heavy 15×15 Gaussian blur left OCR accuracy at a perfect 1.0000. Tesseract's real recognition pipeline is evidently resilient to this specific kind of degradation for this clean, high-contrast text.

Real rotation: a sharp, specific cliff¶

rotation= 5 deg: real char-accuracy=1.0000
rotation=10 deg: recognized='ee mnuter Visio' real char-accuracy=0.3333
rotation=20 deg: recognized='' real char-accuracy=0.0000

A real, dramatic collapse between two adjacent real test points: perfect at 5 degrees, nearly unreadable (0.3333) at 10, and completely blank at 20. Rotation is, by a wide margin, this lesson's most damaging real degradation.

Real noise: robust, then a real cliff of its own¶

noise sigma=50: recognized='‘Computer Vision Mastery Path 2026 S :' real char-accuracy=1.0000
noise sigma=80: recognized='' real char-accuracy=0.0000

A real, different failure shape: accuracy holds at a perfect 1.0000 even with visibly messy extra characters appearing around the real correct text (the matching algorithm still finds the real substring intact), then falls completely to 0.0000 once noise crosses a real threshold between sigma=50 and sigma=80.

Real word-level detection, not just a raw string¶

boxes_data = pytesseract.image_to_data(canvas, output_type=pytesseract.Output.DICT)

word='Computer' confidence=96 box=(22,36,164,31)
word='Vision' confidence=96 box=(194,35,98,25)

Real, individually-boxed words, each with its own real confidence score — the actual, structured output a real document-processing pipeline would use, rather than a single opaque string.

How It Actually Works¶

Tesseract's real recognition pipeline is built around detecting character shapes along consistently horizontal (or a few expected) text lines — blur softens edges but preserves the real overall shape and horizontal alignment of each character, which is exactly why this lesson's blur test stayed perfect. Rotation does something blur doesn't: it breaks the real horizontal-line assumption the segmentation step depends on, so even a moderate real rotation (10 degrees) can confuse where one line of text starts and ends, producing the kind of garbled partial output this lesson actually measured rather than a graceful accuracy decline.

Noise's real cliff shape — fine up to a point, then total failure — reflects a different real mechanism: Tesseract's thresholding step can usually still separate real dark text from a noisy light background up to a point, but once noise gets severe enough to genuinely obscure character edges, the whole segmentation step fails at once rather than degrading smoothly character by character.

Common mistakes¶

Assuming OCR degrades gracefully and roughly proportionally to how degraded an image looks to a human eye — this lesson's real rotation and noise results both show sharp cliffs, not smooth declines, and the cliffs sit at real, different, non-obvious thresholds for each kind of degradation.
Preprocessing for "general image quality" instead of the real, specific degradation OCR is sensitive to — this lesson's results suggest a real document pipeline should prioritize deskewing (correcting rotation) far more than denoising or deblurring, since rotation was the most damaging degradation measured here by a wide margin.
Reading only the raw recognized string and ignoring per-word confidence — this lesson's real image_to_data() output gives a structured, per-word confidence score that a raw string alone cannot, useful for flagging specifically which words in a real document might need human review.

Exercise¶

Re-run this lesson's exact real rotation test at every 1-degree increment from 5 to 10 degrees, to find the real, specific angle where character accuracy actually starts collapsing — narrowing down exactly where this lesson's observed cliff between 5° (perfect) and 10° (0.3333) really sits.