TrOCR Base Printed

microsoft/trocr-base-printed is an image-to-text model that performs optical character recognition (OCR) on printed text-line images using a pure Transformer encoder-decoder architecture, without CNN or RNN components.

Architecture and Training

The model uses a Vision Transformer (initialized from BEiT) as the encoder and a text Transformer (initialized from RoBERTa) as the decoder. Images are processed as 16×16 patches with linear embedding and absolute position encoding, then autoregressively decoded into wordpiece tokens. TrOCR is trained end-to-end: first pre-trained on large-scale synthetic data, then fine-tuned on the SROIE dataset for printed text recognition. The approach was introduced in Li et al. (2021), published at AAAI 2023, with 334 million parameters.

Benchmark Performance

On the SROIE dataset, TrOCR-Base achieves an F1 score of 96.34 (96.60 for the Large variant). The model also demonstrates strong results on scene text recognition benchmarks:

On the IAM handwritten dataset, the model achieves a cased character error rate (CER) of 3.42, confirming its capability beyond printed text.

Key Strengths

• End-to-end Transformer architecture eliminates the need for separate CNN, RNN, and language model post-processing steps.
• Pre-training on synthetic data followed by targeted fine-tuning enables strong generalization across printed, handwritten, and scene text domains.
• Competitive benchmark results against state-of-the-art models at the time of publication.