Struct SLANetModel 

pub struct SLANetModel { /* private fields */ }

Pure SLANet model implementation.

Implementations

impl SLANetModel

pub fn new( inference: OrtInfer, normalizer: NormalizeImage, input_shape: InputShape, ) -> SLANetModel

Creates a new SLANet model.

pub fn input_shape(&self) -> &InputShape

Returns the input shape of this model.

pub fn preprocess( &self, images: Vec<ImageBuffer<Rgb<u8>, Vec<u8>>>, ) -> Result<(ArrayBase<OwnedRepr<f32>, Dim<[usize; 4]>>, Vec<[f32; 6]>), OCRError>

Preprocesses images for table structure recognition.

Preprocessing strategy depends on model type:

• SLANeXt (fixed 512×512): ResizeByLong + Padding to square
• SLANet/SLANet_plus (dynamic): ResizeByLong only, no padding

This difference is required because the ONNX-exported SLANet_plus model produces incorrect (truncated) output when given padded square input, while SLANeXt models require fixed 512×512 input.

pub fn infer( &self, batch_tensor: &ArrayBase<OwnedRepr<f32>, Dim<[usize; 4]>>, ) -> Result<(ArrayBase<OwnedRepr<f32>, Dim<[usize; 3]>>, ArrayBase<OwnedRepr<f32>, Dim<[usize; 3]>>), OCRError>

Runs inference on the preprocessed tensor.

Returns dual outputs: structure logits and bbox predictions.

pub fn forward( &self, images: Vec<ImageBuffer<Rgb<u8>, Vec<u8>>>, ) -> Result<SLANetModelOutput, OCRError>

Runs the complete forward pass: preprocess -> infer.

Postprocessing is handled separately by TableStructureDecode.

Model Output Order

The ONNX model outputs are in this order:

• fetch_name_0: bbox predictions [batch, seq, 8]
• fetch_name_1: structure logits [batch, seq, vocab_size]

Trait Implementations

impl Debug for SLANetModel

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.