dart_mupdf_donut 1.0.0

A comprehensive pure Dart PDF library with OCR-free document understanding. Combines PyMuPDF-inspired PDF parsing (text/image extraction, annotations, page manipulation, PDF creation) with a Donut (Sw [...]

1.0.0

Initial release of dart_mupdf_donut — a pure Dart PDF library with OCR-free document understanding.

PDF Module (dart_mupdf)

• PDF parsing engine (cross-reference table, indirect objects, streams)
• Document class — open from file/bytes, save, metadata, TOC, page manipulation
• Page class — text extraction, images, links, annotations, search
• Text extraction: plain text, blocks, words, dict, HTML, XML, XHTML
• Image extraction from PDF pages
• Annotation reading and creation (highlight, underline, strikeout, text, freetext, ink, stamp)
• Table of contents extraction and modification
• PDF manipulation: merge, split, rotate, delete, copy, move pages
• Metadata reading and writing
• Link extraction and creation
• Form field (widget) reading
• Embedded file support
• Page labels
• Encryption detection and password authentication
• PDF creation from scratch
• Cross-reference (xref) object access
• Geometry types: Rect, Point, Matrix, Quad, IRect
• Colorspace support
• Deflate/inflate stream compression
• Drawing API (Shape) for lines, rectangles, circles, curves
• Pixmap — pixel buffer with colorspace conversion and PNG export
• Pure Dart — no native dependencies, works on all platforms

Donut Module (OCR-free Document Understanding)

• Pure Dart implementation of Donut (ECCV 2022)
• Swin Transformer encoder — hierarchical vision transformer with shifted windows, patch embedding, patch merging
• mBART decoder — auto-regressive text decoder with cross-attention to encoder features, KV caching for fast generation
• Tensor class — N-dimensional array with Float32List, matmul, softmax, GELU, broadcasting, sum/mean/max reductions, reshape/permute/transpose
• Neural network layers: Linear, LayerNorm, Embedding, Conv2d, MultiHeadAttention, FeedForward
• DonutTokenizer — SentencePiece BPE tokenizer compatible with HuggingFace format, special token support
• DonutImageUtils — image preprocessing (resize, pad, ImageNet normalization), tensor ↔ image roundtrip
• DonutWeightLoader — load weights from safetensors or JSON format
• DonutConfig — full configuration (base, small, custom, fromJson)
• JSON ↔ Donut token conversion (json2token / token2json)
• Support for CORD-v2 (receipts), RVL-CDIP (classification), DocVQA, SynthDoG tasks
• Random weight initialization for testing (model.randomInit())
• Compatible with HuggingFace pretrained Donut models