Tabelstrukturgenkendelse

Tabelstrukturgenkendelse is a term used to describe the process of identifying and reconstructing the structure of tabular data within documents or images. It focuses on determining the layout of a table, including the location of the table, the arrangement of rows and columns, and the boundaries of individual cells, as well as handling features such as spanning and merged cells. The goal is to convert visual or textual table representations into a structured, machine-readable format.

The task typically comprises two interrelated subtasks: table detection, which locates and isolates tables in a page, and table structure recognition, which interprets the internal organization of the detected table. Structure recognition involves predicting the grid of cells, assigning content to each cell, and understanding headers, captions, and footnotes. Modern approaches often combine computer vision and machine learning, using end-to-end deep learning models or hybrid methods that incorporate rule-based heuristics, layout analysis, and post-processing to produce a usable representation such as a CSV or a JSON table.

Common techniques include region-based object detection, sequence labeling, graph-based representations of tables, and models that reason about row and column relationships. Datasets for benchmarking include publicly available table recognition and table parsing corpora, with metrics evaluating both the accuracy of cell detection and the correctness of the resulting table structure and content.

Applications of tabelstrukturgenkendelse include digitizing printed or scanned documents, extracting data from financial reports, scientific papers, and historical archives, and enabling downstream data analysis, indexing, and searchable archives. Challenges remain with irregular layouts, multi-page or nested tables, merged cells, rotated or degraded images, and multilingual or multi-format documents.