Why my scanned drawing reads badly, and the fixes
The process engineer runs automatic analysis on a set of drawings scanned from paper archives printed in the 1990s. The optical reading engine misses half the toleranced dimensions, mistakes a zero for a capital C and ignores the GD&T feature control frames. The team loses precious time retyping inspection tables that were supposed to be generated automatically.
The physical and digital reasons a scanned drawing reads badly
Understanding why a scanned drawing reads badly means looking at the whole capture chain, from the state of the original paper to the compression format of the PDF produced by the shop scanner. Poor legibility comes from low resolution, aggressive binarisation or geometric distortion from folded sheets.
| Capture parameter | Faulty setting | Recommended setting | Impact on OCR |
|---|---|---|---|
| Image resolution | 150 dpi or lower | 300 dpi monochrome or 8-bit greyscale | Pixelated tolerance figures |
| File format | Heavily compressed JPEG | LZW TIFF or lossless PDF | Background noise and artefacts around the text |
| Binarisation | Fixed 50% threshold | Dynamic adaptive binarisation | Fine dimension lines wiped out |
| Physical alignment | Skew above 1.5° | Automatic deskew | Text blocks wrongly grouped |
A poor setting on just one of these four pillars is enough to drop the recognition rate of technical text below the usable threshold.
Insufficient resolution and undersampled fine text
The main reason a scanned drawing fails to read is a resolution below 300 dpi (dots per inch). At 150 dpi, subscript tolerance figures (often no taller than 1.5 mm on paper) are coded on only 5 to 7 pixels of height.
At that level of pixelation, the loop of an 8 merges into a 0 and the top bar of a 5 vanishes, turning the requirement into an unreadable 3.
To tell technical characters apart reliably, capture must be done at a true 300 dpi. Scanning at 600 dpi brings no measurable gain for engineering OCR, while making the files four times heavier.
Lossy compression and JPEG artefacts
Saving scans as JPEG with a high compression ratio creates 8x8 pixel blocks of artefacts around high-contrast lines.
The recognition engine takes these clusters of noise for dots, commas or leader lines. A 12,5 dimension polluted by compression noise ends up read as 125 or 12.5.0, which makes automatic conversion impossible.
Always prefer a PDF export with lossless LZW or CCITT Group 4 compression for monochrome documents.
Looking at the automatic dimension detection on PDF tools shows why a clean pixel structure matters.
Geometric skew and fold shadows
An A1 print folded in four keeps deep fold marks. When it goes through the shop’s sheet-feed scanner, the fold creates a dark horizontal line across the dimensions.
If the sheet feeds in at an angle of more than 1.5 degrees, the line segmentation algorithms fail to group the nominal diameter with its tolerance.
Automatic deskew must be switched on in the scanner driver, to realign the title blocks and dimension lines with the X and Y axes of the image.
Example: improving a scanned archive drawing
A metrology laboratory has to scan an assembly drawing of a hydraulic valve body dimensioned ⌀12 h7. The first scan, done at 150 dpi as a JPEG by the shop’s automatic scanner, produces a blurred image skewed by 2.3 degrees.
The analysis engine reads the dimension as 012 h1 instead of ⌀12 h7, triggering a conversion error alert.
- Take the original paper document back to the scanning station.
- Clean the grease marks off the title block with a drafting eraser.
- Set the scanner to 300 dpi, 8-bit greyscale.
- Turn on the automatic deskew filter.
- Export as an uncompressed PDF.
Result after the new capture: the ⌀12 h7 dimension is captured in full, with its exact tolerance characters. The extraction rate of dimensional characteristics is back to 100% across the whole document.
Shop-floor pitfalls when scanning drawings
The first pitfall is scanning in pure binary mode (1-bit black and white) with a fixed binarisation threshold. If the dimension line is slightly faded on the original paper, binarisation cuts the leader line and separates the nominal value from the diameter symbol. Scanning in greyscale keeps the shading information and allows more effective dynamic filtering.
The second comes from shop rubber stamps (“APPROVED FOR MANUFACTURE” or similar) applied in red or blue ink over the dimensions. When converted to black and white, the stamp mark hides the figures underneath completely. Scanning in colour with digital removal of the secondary ink layer (colour drop-out) brings the hidden dimensions back.
The third concerns crumpled folds that create false dimension arrows. A diagonal fold across text looks like a section line or an extra leader line, which confuses vector structure analysers.
To process these scanned documents efficiently without losing data, CoteCote runs 100% locally on Windows to keep your drawings fully confidential, with processing from 200 ms per page on vector PDFs and up to 98% reading accuracy on scanned documents.