Overview
Academic publications across clinical and engineering formats, including IEEE and ICML, present diverse layout types containing both vector math charts and raster images. Traditional layout parsers struggle when text columns sit adjacent to unlabeled charts, or when scanner outputs exhibit rotational skew. This advanced analyzer extracts structured assets entirely in the browser using spatial projection profiling. By evaluating both visual pixel maps and standard text distributions, the extractor isolates target diagrams while leaving page metadata and text columns untouched.
How to Use
To analyze a clinical paper or document, drag and drop your PDF into the upload zone or click to open your system directory. The engine maps textual coordinates, calculates document columns, and segments diagram assets page by page. Use the specific tabs to browse cleanly cropped figures, automatically reconstructed LaTeX table matrices (copiable as clean CSV grids), and raw paper text. If a custom diagram falls outside standard caption anchors, select the page in the Snip Tool and drag a manual bounding box to cleanly extract and download any region.
Technical Details
The core extraction engine utilizes an innovative Text-Density Radar Mapping paradigm implemented on top of PDF.js. To eliminate edge-trimming failures and avoid collecting plain paragraph elements, the extractor evaluates the physical surface area occupied by text bounding boxes inside any prospective visual block. If the text ratio exceeds a strict 35% threshold, the crop is classified as standard text and rejected. Furthermore, a dynamic rotation/skew optimization algorithm rotates candidate canvas objects through minor angle offsets (ranging from -3 to +3 degrees) to find the orientation maximizing the variance of the horizontal projection profile. Connected-component clustering isolates unlabeled graphs, equations, and diagrams.
Future Directions
The current iteration supports double-column formats, clinical layouts, and scan deskewing configurations. Upcoming releases will feature client-side WebAssembly OCR implementations designed to reconstruct text inside scanned documents that lack digital fonts, allowing clinicians and researchers to apply structural projection mapping to legacy medical records and historically archived papers.