TrOCR Small Printed

microsoft/trocr-small-printed is an image-to-text model that performs optical character recognition (OCR) on printed text lines, fine-tuned on the SROIE dataset.

Architecture and Approach

The model uses an encoder-decoder Transformer architecture. The image encoder is initialized from DeiT weights and processes input images as a sequence of 16x16 patches with absolute position embeddings. The text decoder, initialized from UniLM, autoregressively generates wordpiece-level tokens. This end-to-end design eliminates the need for separate CNN, RNN, and language model components, as demonstrated in the TrOCR paper (Li et al., AAAI 2023).

Benchmark Performance on SROIE

The SROIE (ICDAR 2019 Scanned Receipts OCR and Information Extraction) benchmark evaluates OCR on challenging receipt images with poor print quality, low resolution, folded documents, and privacy-blurred fields. The following table compares TrOCR variants on SROIE F1 score, as reported in the official GitHub repository:

The small variant thus achieves competitive accuracy with significantly fewer parameters, making it a lightweight option for printed text recognition.

Capabilities and Context

The model is designed for single text-line images. According to the paper abstract, the TrOCR family outperforms prior state-of-the-art models on printed, handwritten, and scene text recognition tasks. This particular checkpoint is specialized for printed text, leveraging the SROIE training set to handle real-world receipt variability.