Module vision 

Optional vision layer: raster plus ONNX runner (#636). Off by default. Optional vision layer: page raster plus ONNX runner (#636).

Behind the vision cargo feature (off by default). No trained model ships; a separate team plugs models in later. This module is infrastructure only:

• rasterize_page_with_dpi renders a page through the engine’s existing renderer (crate::PdfDocument::render_page) and converts the RGBA output to grayscale or RGB bytes. No new renderer.
• VisionRunner loads an ONNX model from bytes or a path with tract-onnx (MIT OR Apache-2.0) and runs it on a VisionRaster or on a crop named by a bbox in page points, returning the output tensor as VisionOutput.

Input tensors are NCHW f32 in 0.0..=1.0 (bytes divided by 255). Output is the model’s first output, which must be f32.

Structs

VisionOutput

The output tensor of one model run: shape plus row-major f32 data.

VisionRaster

A page (or crop) rasterized for vision input, row-major.

VisionRunner

An ONNX model loaded from bytes or a path, runnable on a VisionRaster.

Enums

VisionError

Errors from the vision layer.

VisionPixelFormat

Pixel layout of a VisionRaster.

Constants

DEFAULT_VISION_DPI

Default rasterization resolution for vision input, in dots per inch.

Functions

rasterize_page

Rasterize a page to a grayscale or RGB buffer at DEFAULT_VISION_DPI.

rasterize_page_with_dpi

Rasterize a page to a grayscale or RGB buffer at the given dpi.

Type Aliases

VisionResult

Convenience alias for vision operations.