pix2text-mfr

Recognizes and extracts mathematical formulas from document images using a vision-encoder-decoder architecture that combines a visual encoder (processes image patches) with a sequence decoder that outputs LaTeX representations. The model is trained to handle handwritten and printed mathematical notation, converting visual mathematical content directly into machine-readable LaTeX strings without intermediate OCR steps.

Performs optical character recognition on printed text in document images using the same vision-encoder-decoder backbone, converting visual text content into machine-readable strings. The encoder processes image patches through a convolutional or transformer-based visual feature extractor, while the decoder generates character sequences autoregressively, handling multi-line text and variable document layouts.

Provides ONNX-format model export enabling efficient batch inference on CPU or specialized hardware without PyTorch dependencies. The model can be loaded via ONNX Runtime, which applies graph optimization, operator fusion, and quantization-aware execution paths, reducing latency and memory footprint for production deployments. Supports batching multiple images in a single inference call for throughput optimization.

Recognizes and extracts text from documents in multiple languages using a language-agnostic vision-encoder-decoder trained on diverse multilingual corpora. The visual encoder is language-independent (processes image features), while the decoder is trained to generate character sequences in multiple languages, handling script variations (Latin, Cyrillic, CJK, Arabic, etc.) without language-specific preprocessing.

Implements a two-stage neural architecture where a vision encoder (CNN or Vision Transformer) extracts spatial features from document images, and a sequence decoder (RNN or Transformer) generates output text autoregressively. The encoder processes variable-size images by patching or resizing, producing a fixed-size feature representation; the decoder consumes this representation and generates tokens sequentially, with attention mechanisms enabling focus on relevant image regions during generation.

Generates valid LaTeX code directly from mathematical formula images, producing strings that can be compiled by LaTeX engines without post-processing. The decoder is trained on LaTeX syntax and mathematical notation conventions, learning to generate properly balanced braces, escaped special characters, and valid command sequences. Output can be directly embedded in LaTeX documents or mathematical typesetting systems.