Struct OcrLite 

pub struct OcrLite { /* private fields */ }

Implementations

impl OcrLite

pub fn new() -> Self

pub fn set_doc_orientation_model(&mut self, clf: DocOrientationClassifier)

Carica il classificatore orientamento pagina (PP-LCNet_x1_0_doc_ori_onnx). Necessario per OcrOptions { use_doc_orientation: true }.

pub fn init_models( &mut self, det_path: &str, cls_path: &str, rec_path: &str, num_thread: usize, ) -> Result<(), OcrError>

pub fn init_models_with_dict( &mut self, det_path: &str, cls_path: &str, rec_path: &str, dict_path: &str, num_thread: usize, ) -> Result<(), OcrError>

pub fn init_models_no_angle( &mut self, det_path: &str, rec_path: &str, dict_path: &str, num_thread: usize, ) -> Result<(), OcrError>

Variante senza angle-net: init solo det + rec, salta il cls model. Usare con do_angle: false (se do_angle=true l’inferenza fallisce con SessionNotInitialized). Utile nei test dove il cls ONNX non è disponibile in cache.

pub fn init_models_with_dict_and_builder( &mut self, det_path: &str, cls_path: &str, rec_path: &str, dict_path: &str, num_thread: usize, builder_fn: Option<fn(SessionBuilder) -> Result<SessionBuilder, Error>>, ) -> Result<(), OcrError>

Variante di init_models_with_dict che accetta un builder_fn custom da applicare a tutti e 3 i Session::builder() (det+cls+rec). Usato dai consumer (Edge) per registrare execution provider hardware-accelerated (QNN-HTP / DirectML / CoreML / CUDA / XNNPACK) invece del default CPU-only del path standard.

pub fn init_models_custom( &mut self, det_path: &str, cls_path: &str, rec_path: &str, builder_fn: fn(SessionBuilder) -> Result<SessionBuilder, Error>, ) -> Result<(), OcrError>

pub fn init_models_custom_with_dict( &mut self, det_path: &str, cls_path: &str, rec_path: &str, dict_path: &str, builder_fn: fn(SessionBuilder) -> Result<SessionBuilder, Error>, ) -> Result<(), OcrError>

pub fn init_models_from_memory( &mut self, det_bytes: &[u8], cls_bytes: &[u8], rec_bytes: &[u8], num_thread: usize, ) -> Result<(), OcrError>

pub fn init_models_from_memory_custom( &mut self, det_bytes: &[u8], cls_bytes: &[u8], rec_bytes: &[u8], builder_fn: fn(SessionBuilder) -> Result<SessionBuilder, Error>, ) -> Result<(), OcrError>

pub fn detect( &mut self, img_src: &RgbImage, padding: u32, max_side_len: u32, box_score_thresh: f32, box_thresh: f32, un_clip_ratio: f32, do_angle: bool, most_angle: bool, ) -> Result<OcrResult, OcrError>

检测图片

Arguments

• &self (undefined) - Describe this parameter.
• img_src (&image) - 图片
• padding (u32) - 变换图片时添加边框的宽度（提高检测效果）
• max_side_len (u32) - 变换图片后图片宽和高保留的最大边长（超出该尺寸的图片将缩小）
• box_score_thresh (f32) - 检测存在文本的区域的分值阈值
• do_angle (bool) - 是否进行角度检测

pub fn detect_angle_rollback( &mut self, img_src: &RgbImage, padding: u32, max_side_len: u32, box_score_thresh: f32, box_thresh: f32, un_clip_ratio: f32, do_angle: bool, most_angle: bool, angle_rollback_threshold: f32, ) -> Result<OcrResult, OcrError>

支持角度回滚的检测图片 在 do_angle 为 true 时生效，如果图片经过了角度纠正，但识别效果过差，则取消角度纠正

Arguments

• &self (undefined) - Describe this parameter.
• img_src (&image) - 图片
• padding (u32) - 变换图片时添加的边框的宽度（提高检测效果）
• max_side_len (u32) - 变换图片后图片宽和高保留的最大边长（超出该尺寸的图片将缩小）
• box_score_thresh (f32) - 检测存在文本的区域的分值阈值
• do_angle (bool) - 是否进行角度检测
• angle_rollback_threshold (f32) - 角度回滚的阈值，如果识别到的文字得分低于该值（或等于 NaN），则取消角度回滚

pub fn detect_from_path( &mut self, img_path: &str, padding: u32, max_side_len: u32, box_score_thresh: f32, box_thresh: f32, un_clip_ratio: f32, do_angle: bool, most_angle: bool, ) -> Result<OcrResult, OcrError>

pub fn detect_with_layout( &mut self, image: &RgbImage, layout: &mut LayoutAnalyzer, padding: u32, max_side_len: u32, box_score_thresh: f32, box_thresh: f32, un_clip_ratio: f32, do_angle: bool, most_angle: bool, options: OcrOptions, ) -> Result<LayoutAwareResult, OcrError>

Pipeline layout-aware completa: cls (per-line) → layout → ppocr (det+cls+rec) → associazione text-line ↔ layout-box → orphan recovery (nearest-neighbor) → sort per reading-order.

Vincolo utente: le text-line di OCR sono fonte di verità per il testo (non si filtrano per layout, evitando i buchi quando il layout omette regioni). Il layout serve per ordering + classifica semantica + associazione spaziale.

Ordering:

1. Primary: LayoutBox.reading_order (-1 va in fondo).
2. Secondary: y-position del centroide del text-block dentro lo stesso box (top-to-bottom).
3. Tertiary: x-position del centroide (left-to-right) — utile per box che contengono testo affiancato.

Orphan recovery: text-line con centroide fuori da TUTTI i layout-box → assigned al box più vicino (centroid distance). Se non ci sono layout-box → tutto torna come orphan ordinato per y/x (fallback graceful per pagine senza layout detection).

pub fn detect_with_options( &mut self, img_src: &RgbImage, padding: u32, max_side_len: u32, box_score_thresh: f32, box_thresh: f32, un_clip_ratio: f32, do_angle: bool, most_angle: bool, options: OcrOptions, ) -> Result<OcrResult, OcrError>

Variante “ricca” di Self::detect con OcrOptions.

I flag use_seal, use_formula, use_chart, use_doc_orientation e use_doc_unwarping sono riservati e non ancora implementati: se impostati a true viene emesso un warning su eprintln! e l’esecuzione prosegue senza lo stage corrispondente. Questo garantisce che i caller esistenti non si rompano quando gli stage verranno aggiunti in futuro.

Trait Implementations

impl Debug for OcrLite

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

Formats the value using the given formatter.

impl Default for OcrLite

fn default() -> Self

Returns the “default value” for a type.