Table extraction from images in Computer Vision - Model Pipeline Trace

Model Pipeline - Table extraction from images

This pipeline extracts tables from images by detecting table regions, recognizing lines and cells, and then converting them into structured data like CSV or JSON.

Data Flow - 6 Stages

1Input Image

1 image (e.g., 1024 x 768 pixels, 3 color channels)→Raw image of a document page containing tables→1 image (1024 x 768 x 3)

Photo of a printed page with a table of sales data

↓

2Preprocessing

1 image (1024 x 768 x 3)→Resize, grayscale conversion, noise reduction→1 image (512 x 384 x 1)

Grayscale image with reduced noise and normalized brightness

↓

3Table Detection

1 image (512 x 384 x 1)→Detect bounding boxes around tables using CNN→List of bounding boxes (e.g., 3 boxes)

Detected boxes: [{"x":50,"y":100,"w":400,"h":200}, {"x":500,"y":150,"w":300,"h":180}, {"x":100,"y":400,"w":350,"h":150}]

↓

4Cell Segmentation

Each table image crop (variable size)→Detect rows and columns lines to segment cells→Grid of cells (e.g., 10 rows x 5 columns)

Table cropped and segmented into 50 cells

↓

5Text Recognition (OCR)

Each cell image (small cropped region)→Recognize text inside each cell using OCR→Text strings for each cell

Cell texts: [['Date', 'Product', 'Price'], ['2024-01-01', 'Pen', '$1.20'], ...]

↓

6Structured Output

Text strings for all cells→Combine cell texts into structured table format (CSV/JSON)→Structured table data (e.g., JSON with rows and columns)

{"rows": [{"Date": "2024-01-01", "Product": "Pen", "Price": "$1.20"}, ...]}

Training Trace - Epoch by Epoch

Loss
1.2 |*       
0.8 | **     
0.5 |  ***   
0.3 |    ****
0.2 |     *****
     ----------------
      1  3  5  7  10 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Initial training with high loss and low accuracy on table detection
3	0.8	0.65	Model starts to detect tables more accurately
5	0.5	0.80	Improved detection and segmentation of table cells
7	0.3	0.90	High accuracy in detecting tables and segmenting cells
10	0.2	0.94	Model converged with low loss and high accuracy

Prediction Trace - 5 Layers

Layer 1: Input Image

Layer 2: Table Detection

Layer 3: Cell Segmentation

Layer 4: Text Recognition (OCR)

Layer 5: Structured Output

Model Quiz - 3 Questions

What is the main purpose of the 'Cell Segmentation' stage?

A To split the detected table into individual cells

B To detect the location of tables in the image

C To convert the image to grayscale

D To recognize text inside each cell

Key Insight

This visualization shows how a model learns to detect tables and segment them into cells, then uses OCR to extract text. The training improves detection accuracy and reduces errors, enabling structured data extraction from images.