Detect Document Layout (PP DocLayout)

@paddlepaddle1 VersionDetect

Detects layout regions on each page Image with PP-DocLayout (PaddleX) and emits one detection per region as geometry-only [Rectangle<Double>] or [BoundingBox] carrying the model's 23-category layout class id and confidence, fixed at deploy by the consumer.

How it fits

Typical backends

A page read from a file is scanned for layout regions and the region boxes are published as JSON for a structure-aware consumer.

Each detected region is cropped from the page for region-restricted OCR downstream.

A live webcam page stream is scanned and the region boxes are published as JSON.

Caveats

Hard constraintuse_gpu set true selects the first GPU at startup and requires an actual GPU; there is no silent CPU fallback for this knob.
I/O contractExactly one arm is emitted, fixed at deploy by the consumer: the [BoundingBox] arm carries the model's own layout label id from the 23-category map, while the [Rectangle<Double>] arm carries no class.
Compatibilitypaddlex_models and paddlex_revision are cache-seeding hints; the list must contain layout_model_name, otherwise the missing model directory is downloaded at first run.
Parameter interactionthreshold drives recall directly; lower values keep more regions but admit noise. layout_nms, layout_unclip_ratio, and layout_merge_bboxes_mode shape overlapping and nested detections.
State lifetimeEvery config key is captured once at startup; there are no hot-swappable parameters.

Versions

d82babb2latestdefault
Automated release
6/28/2026