OmniDoc# how to use it in practice: the processing path of scanned form
This article sorts out several common ideas for table recognition, explains what materials each method is suitable for, and its respective limitations. It helps you determine which path to take for tables in scanned documents, pictures, and digital PDFs, and explains the position of OCR among them.
#What are the table recognition methods: common ideas and applicable scenarios
Key points of this article
- Table recognition ideas are roughly divided into two categories: reading the internal structure of the document, and OCR first and then restoring the table. Whether the material is a digital document or an image determines the starting point.
- Complete table recognition usually includes three steps: detection, structure recognition and content recognition. Errors in any step will lead to data misalignment.
- Tables with borders are suitable for line detection and rule analysis, while tables without borders rely more on layout alignment and text position inference.
- The text in the scanned table needs to be recognized first. After the recognition is completed, the merged cells, headers and numbers still need to be checked.
Many people search for "What are the table recognition methods?" They usually have a scanned copy, a screenshot, or a PDF, and want to turn the table inside into Excel data that can be calculated and filtered. The problem is that table recognition is not a fixed action. There are several types of ideas behind it, and each of them is suitable for different materials. If you choose the wrong idea, the common results are misaligned columns, missing merged cells, or numbers being recognized as words with similar shapes.
This article only talks about the classification of methods and applicable scenarios at the conceptual level, without going into step-by-step operations or comparing specific tools. After reading this, you will be able to judge which category the material you have belongs to, know which way to look for solutions, and where OCR is in the entire process.
The core of table recognition is to restore "things that look like tables" to "data with rows, columns, and cells." Below are answers to frequently asked questions.
How many methods are there for table identification? One sentence conclusion
The form recognition method can be understood from two dimensions: input material and technical ideas, with three to four categories being common. The key to selection is to first confirm whether the material is a digital document or an image, and then check whether the table has borders and whether the structure is complex.
| your situation | Recommended practices | reason |
|---|---|---|
| Word or text PDF, the table itself is text | Prioritize reading the internal structure of the document and parsing it directly | The cell information already exists and the error is small. |
| Scans, photos, screenshots | First OCR recognize the text, then restore the table structure | There are only pixels in the image, you must first recognize the text |
| A neat form with clear borders | Line detection plus rule analysis | Frame lines can directly locate rows and columns, and the logic is clear |
| No frame lines, tables arranged by alignment | Layout analysis and whitespace alignment | There are no lines to rely on, and columns can only be inferred from text positions. |
| Many merged cells and complex layout | Structure identification model plus manual verification | Can handle irregular structures, but output still needs to be reviewed |
What steps does table recognition consist of?
Most form identification processes can be broken down into three steps. Understanding these three steps can help you determine where the problem lies.
Detection: Find where the table is
The detection is responsible for encircling table areas on the page, excluding text, images, headers and footers. If the detection is wrong, the common phenomenon is that only half of the table is recognized, or adjacent paragraphs are mistakenly recognized as table content.
Structure: Determine how many rows and columns there are
Structural recognition requires determining how many rows and columns a table has, which cells are merged, and how many layers the table header has. This step determines the output skeleton. If the structure is wrong, no matter how accurate the subsequent text is, it will be filled in the wrong position.
Content: Read each cell
Content recognition is responsible for reading the text and numbers in each cell. Scans rely on OCR, while digital documents directly read existing text. Errors in this step mainly include misrecognition of numbers, symbols and small font content, so important data needs to be checked item by item.
How do several common ideas work?
Based on lines and rules
This method first finds the horizontal and vertical lines on the page, uses the intersections of the lines to divide the cells, and then puts the text into the corresponding grids according to the rules. Its advantage is that its logic is intuitive and its effect is stable on regular financial statements and transcripts. The limitations are also obvious: when the table has no frame lines, or the frame lines are intermittent and the color is very light, the lines are unreliable.
Based on layout and alignment
For tables without borders, the horizontal position of text and line spacing are often used to infer columns. For example, numbers in the same column are usually right-aligned or left-aligned, and the starting positions of the text are similar, so they can be grouped into one column. This method is suitable for three-line tables in papers and parameter tables in manuals. However, when encountering tables with uneven column widths and many text breaks, it is easy to split a cell into two rows.
Structure recognition based on deep learning
This type of method uses a trained model to directly determine the table structure, including row-column relationships and merged cells. It is more adaptable to complex layouts and irregular headers, but the results may still be wrong, especially on low-resolution images or handwritten content. When using it, think of it as "generating a first draft" rather than a "finished product."
Based on OCR plus post-processing
The common route for scanned documents and pictures is: first use OCR to turn the page into text with location information, and then use the previous methods to put the text into a table structure. OCR solves the problem of "recognizing characters", while table restoration solves the problem of "placing in the correct position". Both are indispensable. What you get by doing OCR alone is usually a bunch of text arranged in rows, not a usable table.
Common misunderstandings and areas of confusion
- Recognizing text does not mean restoring the table. Many outputs only arrange the cell contents in reading order, and the row-column relationship is lost, making them unable to be used directly for sorting and calculation.
- Wired meters and wireless meters cannot use the same set of rules. When applying border rules to a table without borders, a common result is that columns are chopped up or merged incorrectly.
- Scans and digital PDFs have different processing paths. PDFs that can select text usually do not require OCR and can be read directly; most of the PDFs that cannot be selected are images and need to be recognized first.
- Merge cells are prone to errors. Table headers and group titles that span rows and columns are often the focus of manual verification.
📊 Core comparison
| Train of thought | Suitable material | A situation at which one is good at | An area prone to problems. |
|---|---|---|---|
| Read the internal structure of the document | Word, literal PDF | The cell boundary is clear and the content is accurate. | Cannot be used when the material itself is a picture |
| Line detection plus rules | A scanned piece or screenshot with a frame. | A regular financial statement, statistical table | The frame lines are intermittent, the color is light, and the lines are redundant. |
| Align the layout with whitespace | Essay table and parameter table without rimline | A table with well-aligned columns | Uneven column width and many line wrapping in the cell |
| Structure recognition model | Tables with complex layouts and multiple merges | Irregular header, across rows and columns | Low-definition pictures, handwritten content |
| OCR plus table restore | All kinds of scanned pieces and photos | Turn an image into editable content | Misrecognition of numbers and symbols, need to be checked |
If you have a scanned piece or picture in your hand, you can follow the order below. The PDF editor of the OmniDoc Desktop client (download.omnidoc.top) includes the OCR feature, which can handle the scanned version of PDF and support export to Word and Excel.
- Judge the type of material first: try to select a paragraph of text in PDF, but the one that is not selected is usually the scanned piece, so you need to do OCR first.
two. Run OCR on the scanned version of PDF to make the text on the page selectable and editable.
- Check that the table area is complete and that the header and merged cells are identified in the correct location.
- Export to an Excel table, and then check the numbers, decimal points, minus signs and units item by item.
- Make a second confirmation of the key data, such as line by line with the original, and then use it for calculation or submission.
If you need to "translate the contents of the table" instead of exporting the data, you can also upload a scanned version of PDF on omnidoc.top, choose bilingual control or translation only, and also need to check the table structure and proper nouns after download. Both machine translation and OCR can make mistakes. Please review the important contents item by item.
# frequently asked questions (FAQ)
1. There are several methods of form recognition, and which are the common ones?
According to the input material, there are two common types: reading the internal structure of the digital document, and OCR the image first and then restore the table. According to the technical thinking, the common ones are line detection plus rules, layout alignment, structure recognition model and OCR plus post-processing. In practice, it is often a combination of several ideas.
2. How to change the Excel table in the picture into editable data?
First make sure the picture is clear and the table is complete, then use OCR to identify the text, and then restore the row and column structure. Check the numbers and headers item by item after the identification is completed. If the picture is tilted or blurred, it is recommended to re-shoot or scan first, the recognition quality will be more stable.
3. What is the difference between a form with and without a frame line?
A table with a frame can rely on lines to locate cells with clear logic. A table without a frame can only infer columns based on the position and alignment of the text, which has a higher probability of error and usually requires more manual checking.
4. What if the number is wrong after identification?
First go back to the original to confirm whether it is a recognition error, such as 0 and O, 1 and l, decimal point omission and other common cases. When regular errors are found, you can replace them in batches and check them again. The key figures must be compared with the original item by item.
5. What should I look at when choosing a free form recognition tool?
You can focus on several aspects: file size restrictions, whether the output is watermarked, privacy instructions for uploading files, whether the table structure and merged cells can be retained, and whether the export format is suitable for your subsequent use. These can better reflect the actual effect than simply watching "free".
# Summary
The method of form recognition depends on the material in your hand and the appearance of the form itself. The structure of digital documents is read first, and the scanned parts are OCR first and then the tables are restored. The tables with and without frame lines have their own corresponding ideas. After sorting out the concept, then look at the specific steps, the judgment will be faster and there will be fewer mistakes.
