Image preprocessing is the set of operations that clean and standardize a document scan before optical character recognition reads it. In commercial real estate, it deskews, denoises, binarizes, and rescales a photographed or faxed lease so the OCR engine sees crisp black text on white. It runs first, and everything downstream depends on it.
How Does Image Preprocessing Work?
Image preprocessing works by running a scan through a fixed chain of corrective steps before any character is read. Each step targets one defect: deskewing rotates crooked pages to horizontal, binarization forces every pixel to black or white, denoising removes speckle, and rescaling to 300 DPI or higher gives the OCR engine enough resolution to resolve small text.
The order matters. Binarization is usually applied after deskewing so the threshold is computed on aligned text, and denoising follows binarization to strip the salt-and-pepper artifacts that thresholding can introduce. Adaptive binarization, which sets a local threshold per pixel neighborhood, beats a single global threshold on documents with uneven lighting, a common condition in phone-photographed leases.
Step | Defect it corrects | CRE example |
Deskew | Crooked page rotation | A lease scanned at a 5 degree angle |
Binarization | Color and gray backgrounds | A shaded rent schedule table |
Denoising | Speckle and scan artifacts | A faxed estoppel with grain |
Rescaling to 300+ DPI | Low resolution | A 150 DPI mobile capture |
Contrast enhancement | Faint or gray text | An aged, photocopied amendment |
Why Image Preprocessing Matters
Image preprocessing matters because OCR accuracy is capped by input quality, and most CRE documents arrive degraded. Scanning at 300 DPI, removing noise, binarizing, and correcting skew can add 10 to 15 percentage points to raw OCR output, per Docsumo. A 150 DPI mobile photo maxes out around 80 to 85% no matter what runs downstream.
The failure mode is quiet. Skew alone is punishing: a document rotated 5 degrees loses 10 to 15% accuracy, and at 15 degrees the loss reaches 30 to 40%, per DocParser research cited by Docsumo. Those errors do not announce themselves. They land inside a base rent figure or a commencement date, pass through extraction, and price into a deal. Preprocessing is the cheapest accuracy gain in the pipeline because it fixes the input rather than correcting the output.
Example
Image preprocessing is easiest to see as a before-and-after on one page. An analyst receives a 150 DPI phone photo of a lease page skewed 5 degrees with a noisy gray background. The table below applies published preprocessing lifts in sequence to a raw baseline, using the DocParser and Docsumo figures.
Step applied | Reported effect | Running character accuracy |
Raw 150 DPI skewed scan | Baseline | 82% |
Rescale to 300 DPI | Recovers 150 DPI penalty | 88% |
Deskew from 5 degrees | Recovers 5 to 6 points | 93% |
Denoise gray background | Adds up to 15 points, capped here | 96% |
Starting at an 82% baseline on the degraded capture, rescaling recovers the resolution penalty, deskewing recovers the rotation loss, and denoising cleans the background, landing near 96% character accuracy. Each lift is drawn from the cited ranges, not invented, and the running total is capped so no step pushes past a realistic ceiling. The point stands: the same OCR engine reads the same page far better once the image is clean.
Variations and Edge Cases
Image preprocessing changes by document condition, and over-processing can hurt. Aggressive binarization thins already-faint text into gaps, and heavy denoising can erase thin table rules that table extraction relies on. The variants below show where preprocessing helps most and where it backfires.
Condition | Preprocessing treatment |
Clean digital PDF | Little or none; text layer already present |
Faxed or photocopied lease | Full chain: deskew, binarize, denoise |
Mobile phone photo | Add dewarping for perspective distortion |
Faint or low-contrast text | Contrast enhancement, gentle binarization |
Dense table pages | Denoise gently to preserve grid lines |
Image Preprocessing vs OCR
Image preprocessing is often confused with OCR, but preprocessing prepares the image and OCR reads it. Image preprocessing operates on pixels: it deskews, binarizes, and denoises so the page is legible. Optical character recognition then converts that cleaned image into machine-readable characters. One improves the input; the other produces the output.
The relationship is sequential and dependent. OCR never sees the original scan directly in a well-built pipeline; it sees the preprocessed version. This is why two systems using the same OCR engine can report very different accuracy: the difference is often the preprocessing chain in front of it, not the recognition model itself.
Frequently Asked Questions
What is image preprocessing in document extraction?Image preprocessing in document extraction is the cleaning of a scanned page before OCR reads it. It deskews crooked pages, binarizes color backgrounds to black and white, removes noise, and rescales low-resolution scans so the recognition engine sees crisp, legible text.
How much does image preprocessing improve OCR accuracy?Scanning at 300 DPI, removing noise, binarizing, and correcting skew can add 10 to 15 percentage points to raw OCR output, per Docsumo. Deskewing alone matters because a page rotated 5 degrees loses 10 to 15% accuracy, and at 15 degrees the loss reaches 30 to 40%.
Is image preprocessing the same as OCR?No. Image preprocessing prepares the image by deskewing, binarizing, and denoising it. OCR then reads that cleaned image and converts it into characters. Preprocessing improves the input; OCR produces the text output that extraction later structures into fields.
Related Terms
Optical Character Recognition
Extraction Accuracy
Document Extraction
Table Extraction
Confidence Score