Abstract
Patent documents express much of their technical content in drawings, yet a typical compound figure packs several sub-figures together with a dense overlay of reference numerals and lead lines, so downstream tasks such as prior art search and figure-text association must first decompose the figure and separate the annotation text from the line-art. Existing methods treat this as a binary “sub-figure versus background” segmentation that discards the text as noise and encodes no structural prior. We propose PatMamba, a Vision Mamba framework built on a VMamba SS2D backbone and a U-Net decoder that jointly predicts three semantic classes, namely background, sub-figure regions and text-label regions. PatMamba also incorporates a grid-aware scan that appends two grid row/column-priority directions to the four cross-scan directions, thereby encoding the compound-figure layout into the selective-scan sequence at negligible parameter cost. On the DeepPatent2 benchmark, the Mamba backbone lifts figure-level sub-figure detection from for the traditional run-length-smoothing baseline to hit@IoU , ahead of the CNN ( ) and Transformer ( ) backbones; the grid-aware scan lifts the pixel-level mIoU to and yields its largest gains on dense compound figures with four or more sub-figures, improving the sub-figure IoU by and points and the Boundary F1 by points. In zero-shot transfer to utility patents, biomedical compound figures, chemical reaction schemes and handwritten circuit diagrams, this structure-aware bias persists on the visually closest domain and decays in a regular gradient with domain distance.