Case studySearce / Rockwell Automation

Table preservation

Keeping complex tables intact during document extraction.

Hands using a laptop displaying a data table — Photo by Bluestonex / Unsplash ↗

I worked on a document intelligence service for Rockwell Automation at Searce. Part of the job was extracting tables from chemical plant documents.

These tables had merged cells and nested headers. Reading the text alone wouldn’t preserve which header belonged to which value. We needed to keep the structure too.

Keep the table intact#

We fine-tuned CascadeTabNet and used image processing to extract the table structure, including those merged cells and nested headers.

The goal was to carry the relationships through extraction. A table is useful because of how its values fit together.

01 / Table extractionPreserve the structure along with the content.

Plant documentsComplex tables & headers
CascadeTabNetFine-tuning & image processing
Table structureMerged cells & nested headers preserved

Put it into a service#

The extraction work became part of the document intelligence service on GCP, using Kubernetes and Cloud Build.

This was one of those projects where a detail in the document format became a substantial part of the engineering work. The table’s layout was part of the information we needed to keep.