OmniDocPrinciple of invoice recognition software: The recognition process from images to fields
Invoice recognition software generally consists of three steps: first process the images, then recognize the text, and finally assign the text to fields such as amount, title number, and tax number according to layout and rules. This article clearly explains what each step is doing, which steps are prone to errors, and how to verify them.
Key Points of This Article
- The core of invoice recognition consists of two steps: first convert the image into text with location information, then assign the text to fields such as amount and tax ID.
- Image clarity, tilt, and shadows directly affect recognition results; processing before shooting is more convenient than correcting errors afterward.
- Recognizing text does not necessarily mean the fields are correct; amount, tax amount, and date need to be verified against the original image.
- General OCR only reads text; to get the fields in the table, it also looks at the layout position and context.
If you have an invoice photo or a scanned PDF and want to enter the amount, tax number, and invoice date into the form, you will most likely first ask: How exactly does invoice recognition software turn images into these fields? What does invoice recognition software work with? Simply put, it refers to the steps it goes through and what problems each step solves.
This article is written for people who do reimbursement and accounting, and also for readers who want to understand why these tools sometimes mistake "0" for "8." After reading, you'll understand how recognition is divided into several stages, how fields are picked out from a large text pool, where errors are likely to occur, and how to verify them.
What is the principle behind invoice recognition software? In short, a conclusion
The principle of invoice recognition software can be summarized in two steps: first, convert the image into text with positions, then assign the text according to the layout and rules into fields such as amount, payable, and tax number. The first step is text recognition, the next is field extraction; only when both steps are done well can the results be usable.
| Your situation | Recommended recipe | The reason |
|---|---|---|
| I just want to turn a whole scanned page into searchable and copyable text | Identify the entire page using general OCR | You only need the text layer, regardless of which number belongs to which field |
| Enter the amount, date, and tax number in the form | First perform OCR, then check the field positions item by item | Field ownership depends on the layout and context; reading the text correctly does not mean the field is correct |
| There are many invoices and messy layouts | Sample verification and a fixed verification process is set up | Different layouts have different field positions, which can easily cause interlocking |
| Photos are blurry, creased, or shadowed | Retake or scan first, then recognize | If preprocessing is not done well, every subsequent step will be affected |
| Strict requirements for the amount | Manually verify the amount, tax amount, and total price and tax | Errors in the numbers are costly, and machine results require human review |
Image Preprocessing
Before recognition, software usually organizes images. Common methods include: converting to grayscale images, removing background noise, straightening the tilted page, and binarizing the image to further separate text from the background. Shadows, wrinkles, and reflections during photography are all treated as distractions at this step.
The effect of this step depends on the original image. A scanned image with intact edges and uniform lighting requires almost no extra processing; A photo of a crooked phone with missing corners is more likely to cause misidentification.
Text Detection and Recognition (OCR)
After preprocessing, the system first identifies which areas in the image contain text, then recognizes the characters block by block. It outputs not only the text itself but also the coordinates of each paragraph on the page. These coordinates are important because determining whether a number is an "amount" or a "tax amount" mainly depends on its position.
A common misconception is that OCR finishes reading and is done. In fact, OCR only answers "What is written on the image," not "Which item is this?"
Field extraction
The problem to solve in field extraction is: which field does this string of numbers belong to? The system typically references several types of clues:
- Keyword positioning: For example, content near titles like "total price and tax," "invoice date," or "buyer's name" often correspond to the same item.
- Layout structure: The relative positions of rows, columns, borders, titles, and values in the table.
- Formatting rules: Composition of tax code characters, date writing, decimal places, etc.
The same field may appear in different positions on different invoices, so relying solely on fixed coordinates can easily become ineffective, requiring contextual judgment. This is also why errors like "amount linked to tax amount" appear in the recognition results.
Common Misunderstandings
- Misconception 1: Recognizing text is equivalent to identifying a field. There is a layer of classification between text and fields; if classification is wrong, the text itself may be completely correct.
- Misconception 2: Scanned images are processed the same way as phone photos. Scanned samples are usually more regular, while photos require correction and light processing first, resulting in a noticeable difference in effect.
- Misconception 3: If recognition is accurate, there's no need to check. For reimbursement and posting, the amount, tax amount, and date should be itemized against the original; do not skip it just because it "looks right."
📊 Core Comparison
| Category | Output what is it | Suitable for the scene |
|---|---|---|
| Universal OCR | All text and its location on the page | Entire scanned pages can be converted to searchable text and archived materials |
| Invoice field recognition | Fields: amount, tax ID, date, etc | Batch entry, reimbursement ledger, and pre-posting organization |
| Manual data entry | Manually read and fill in each item one by one | Scenarios with few invoices and requiring word-for-word confirmation |
The three do not replace each other. General OCR can serve as a prerequisite for field identification, while manual verification is the final hurdle.
How to use it in practice: Use OmniDoc for text recognition on scans
OmniDoc's desktop client comes with a built-in PDF editor that includes OCR functionality to handle scanned PDFs. Below is a general process suitable for converting scanned invoices into text that can be viewed, selected, and exported:
- Prepare a clear scanned copy or PDF converted from a photo, and try to ensure the page is neat and the text is clear.
- Open the PDF in the OmniDoc desktop client and use the OCR feature to identify the scanned page.
- After recognition, try selecting a segment of amount or tax number to confirm that the text can be selected.
- Compare with the original diagram and verify the amount, tax amount, total price/tax amount, and date item by item.
- If you need to organize it into a table, you can export the results as an Excel file and manually check the field ownership within the table.
It should be noted that OCR discussed in this article is a general text recognition method. It helps you read the text in the image, but field classification still requires you to judge and verify based on the layout.
🎯 Which method suits you?
- Just turn the scanned copy into searchable text and perform OCR directly.
- To enter a large number of invoice fields, first perform OCR, then export the results to the table for row-by-row verification.
- There are only a few invoices with strict amount requirements; manually entering them directly against the originals is actually more reliable.
❓ Frequently Asked Questions (FAQ)
1. What can invoice recognition software recognize?
It can usually recognize printed text and numbers visible on the layout, such as invoice code, invoice number, issuance date, names of buyers and sellers, tax number, amount, tax amount, and total price and tax. The specific effect depends on the software and invoice type; handwritten content and parts covered by stamps are usually less effective at recognizing them.
2. Why is the amount identified sometimes wrong?
Common causes include three types: blurry images causing numbers to be mistakenly identified, such as 0 and 8; tilted or shadowed layouts causing incorrect area divisions; When categorizing fields, adjacent numbers are linked to other fields. When verifying, first look at the number itself, then at the section it belongs to.
3. Can invoices taken with a mobile phone be directly recognized?
You can try, but it's recommended to first take a flat, evenly lit photo with intact corners. When shooting conditions are poor, correcting or reshooting first, usually saves more time than making corrections afterward.
4. Do scans and photos need to be used in the same way?
The principle is the same, but the focus is different. Focus on checking the resolution and page alignment of the scanned sample; Photos should first be corrected for tilt and shadows removed before recognition.
📝 Summary
The core principle of invoice recognition software is to first read images into text with positions, then assign the text to the corresponding fields according to the layout and rules. Once you understand these two steps, you'll see why shooting quality and field verification are just as important. Recognition results can speed up entry, but amounts, tax amounts, and dates should still be checked item-by-item against the originals.
