How Local Table Recognition Works

When you look at a table inside an image, your brain does two things effortlessly at the same time: it reads the words, and it understands which words belong in which row and column. You instantly know that "$25.00" belongs to "Product B" under the "Wholesale Price" column.

For a computer, however, converting a flat table screenshot into an editable Microsoft Excel spreadsheet requires solving a complex two-part puzzle:

1. Reading the characters (What does the text say?)
2. Rebuilding the grid structure (Where does each number live in the matrix?)

Our browser-local tool Local Image to Table solves this by combining two specialized neural networks: PaddleOCR for text reading and SLANet+ for table structure prediction. The entire process runs locally inside your web browser.

Here is an easy-to-understand breakdown of how these models work together to turn image pixels into editable spreadsheet cells.

The Furniture Assembly Analogy: How the Two Models Work Together

Think of converting an image table into a spreadsheet like assembling a modular bookshelf:

• Model 1 (SLANet+): The Empty Bookshelf Frame. This model looks at the image and ignores the words for a moment. It only looks for horizontal rows, vertical columns, and multi-column merged headers. It builds the empty structural skeleton.
• Model 2 (PaddleOCR): The Labeled Books. This model scans the image to find every text snippet, recording its exact coordinates (x, y) on the page.
• The Matchmaker (Spatial Alignment Algorithm): The system places each book into its corresponding shelf compartment based on its coordinates. If a book lands inside Box (Row 3, Column 2), that text becomes the value of cell B3.

If either the bookshelf frame or the book labels are misplaced, your table will look misaligned. Understanding this relationship makes troubleshooting simple.

The 3-Step Local Table Extraction Pipeline

Here is what happens in milliseconds inside your browser when you drop an image into the tool:

Step 1: Text Detection & Character Reading (PaddleOCR)

First, the PP-OCRv6 neural network sweeps across your image:

• It detects bounding boxes around text clusters.
• It identifies individual letters, numbers, currency symbols, and minus signs.
• It records the spatial position of each text line on the canvas.

Step 2: Table Structure Recognition (SLANet+)

At the same time, the SLANet+ vision model analyzes the visual layout geometry:

• It detects physical border lines and whitespace gutters between columns.
• It predicts the total count of rows and columns.
• It identifies merged header cells (such as a single "Q1 2026" header spanning both "Units" and "Revenue" columns).

Step 3: Coordinate Matching and Cell Assembly

Finally, our browser-side algorithm matches the recognized text coordinates with the predicted cell boxes:

• Text that sits inside the boundary of Cell (Row 1, Column 1) becomes the table header.
• Multiple lines of text within a single cell (such as a wrapped address description) are joined together cleanly.
• Blank spaces without any text are assigned as empty cells, ensuring neighboring prices do not slide into the wrong column.

The finished grid renders immediately on your screen as an interactive spreadsheet editor.

The Golden Rule: Structure Matters More Than Characters

If an OCR tool misreads the letter e as c, you can fix that typo in two seconds. But if an OCR tool misses an entire column boundary, every single number to the right shifts by one column! That can scramble hundreds of financial values.

That is why keeping the table structure intact is the most important objective of modern table recognition models.