Technology and architecture
OCR Explained: Character Recognition for Reading Real Business Documents, Complete Guide
Korea Deep LearningPublished
OCR (Optical Character Recognition) is a technology that finds the characters a person can read in a document that exists only as an image, such as a scan or a photo, converts them into text a computer can handle, and also returns the position where each character sits.
Why OCR Is Still Difficult
Printed text is a problem already solved.
For clean printed documents, character reading accuracy reached a practical level long ago. That is why many people believe OCR is a finished technology. But when you open a real intake folder, the picture changes.
In one financial institution's intake documents, a black band on the folded top edge swallowed half of a verification mark. Another document lost its margins and part of its body after being copied many times. There are documents photographed lying flat, with wrinkles, shadows and distortion together, and faxes that were retransmitted four times, so that four different date formats are stacked on one line. Add stamps that run over a table, strikethroughs across characters, and handwriting mixed with checkboxes.
People read these documents naturally. What is hard in character recognition is not reading printed type but reading documents in this condition.
Precise Definition of OCR: How It Differs from Text Extraction
Text Extraction and OCR, Three Axes of Difference
First, the target differs. Text extraction takes out character data that is already in a digitally created document. The text is stored in the source file, so you only need to read it out. OCR targets images that do not have that data. What you see is only pixels, so the work begins with finding the characters inside them.
Second, the goal differs. The goal of text extraction is to move the stored strings without loss. The goal of OCR is to find every character a person can read in the image and restore it as text. That is why text extraction has almost no concept of accuracy, while accuracy is the core metric for OCR.
Third, the metrics differ. Text extraction only checks whether anything was missed. OCR looks at character-level accuracy and similarity-based accuracy together. In real validation cases, a financial document set showed high similarity-based accuracy for both printed and handwritten text, and recognition speed was measured on accelerators. There is also a confirmed case for a document set that includes handwriting.
The two methods are not alternatives. In real business, digital originals and scans arrive in the same workflow, so you need a setup that determines the nature of each file and processes it with the right method. If the original already contains text, converting it to an image and recognizing it only lowers the accuracy.
Six Conditions for OCR That Can Be Used on the Job
The Six Conditions for OCR in Production
First, the document condition must be correctable. Skewed scans, rotated pages and low-resolution fax copies need to be corrected before recognition. Check whether skew correction, rotation correction and super-resolution processing are included upstream.
Second, it must withstand interference. Stamps, watermarks, strikethroughs, overlapping writing and shadows hide or distort characters. If you do not verify with real documents that contain these elements, performance after adoption can differ greatly.
Third, it must handle handwriting. Many items on application forms are filled in by hand. Printed-text accuracy alone does not show real performance, so ask for handwriting accuracy separately.
Fourth, it must return the position of each character. A recognized string alone does not allow item extraction or review afterward. Coordinates for each value must come along so that the original can be pointed to on screen.
Fifth, confidence must be provided per item. A single score for the whole document does not show where to look. Low-confidence parts must be marked separately so that review can focus on exceptions.
Sixth, it must handle the throughput. Judge by the concentration of volume at peak times, not the monthly average. Operating cases measured batch processing time and hourly throughput together.
Practical Application of OCR
Put Preprocessing Before Recognition
The fastest way to raise recognition accuracy is not to change the model but to clean up the input image. A real pipeline converts the document format into images, splits and normalizes pages, corrects skew and rotation, removes noise, and only then passes the result to recognition.
Without this front end, no model placed behind it can recover the losses. For work with many fax copies and phone photos in particular, the quality of preprocessing determines most of the overall performance.
Design Character Recognition and Structure Recognition Together
If you only extract characters in order, the reading order breaks down in tables and in multi-column layouts. In a real setup, the layout is analyzed first to separate titles, body text, tables and figure regions and to set the reading order. Then the characters in each region are recognized.
This makes the recognition result lead directly to structured data. It is far more useful for later item extraction and search than a plain list of strings.
Add Post-Processing Rules
Recognition results contain errors that repeat in patterns. Digits and letters get confused, thousands separators are misread, or the same item repeatedly goes wrong in a particular form.
For these errors, post-processing rules are faster than retraining the model. Items with a fixed shape, such as the digits of a business registration number, the format of a date, or the separators in an account number, are mostly caught when checked with rules.
OCR in the Korean Context
Korean business documents have several features of their own.
First, the document formats. Korean word processor documents are widely used, and this format needs separate handling from the step of converting it to images. Korean government forms and tax returns include circled-number item markers and tables with multiple levels.
Second, the notation. If a document that uses a period as the thousands separator is read as is, the digits of the amount shift entirely. Documents printed without the separator in resident registration numbers are also common. Unless this notation is corrected by rules, it becomes a business error as it is.
Third, the intake route. Fax is still used, and scans from branch offices and phone photos are mixed within one business process. This is an area where a business that has repeated training and validation on Korean documents has an advantage.
Frequently asked questions
Is OCR not already a finished technology?
For printed text, it is. But real business documents mix in stamps, handwriting, low-resolution faxes and damaged scans, and product differences widen considerably here.
What accuracy can I expect?
It depends on the document set. In financial document validation, a case showed high similarity-based accuracy for both printed and handwritten text, and a case with handwritten document sets was also confirmed.
Can it read handwriting?
Yes. However, for items with large variation in handwriting, it is safer to also set validation rules. When reviewing adoption, check handwriting accuracy separately from printed-text accuracy.
How fast is it?
A case measured processing speed per page on an accelerator. Batch processing confirmed hourly throughput. Results vary with the document condition and the hardware.
Can I use the recognition results directly in the business?
Not directly. Recognition results are strings of characters, so they need a step that links each value to its item and rules that verify the format afterward.
Do PDFs that contain text also need recognition?
No. If the original has character data, extracting it directly is more accurate. A setup that determines the nature of each file and splits the processing method accordingly is best.
Can it be used in a network-separated environment?
Yes. With the on-premise build installed on the customer's internal server, it works without external communication.
Send one real intake document and we will return the recognition results as they are.
Contact us





