TrOCR Small Handwritten

TrOCR-small-handwritten is an image-to-text model that performs optical character recognition (OCR) on handwritten text-line images using a Transformer-based encoder-decoder architecture. The encoder, initialized from DeiT, processes images as sequences of 16x16 patches with absolute position embeddings, while the decoder, initialized from UniLM, autoregressively generates wordpiece-level text tokens. This end-to-end approach eliminates the need for separate CNN-based image understanding, RNN-based text generation, or post-processing language models. The model is fine-tuned on the IAM Handwriting Database and achieves a Cased Character Error Rate (CER) of 4.22 on the IAM test set. It has 62 million parameters and was trained on 384-pixel input images. The underlying TrOCR architecture, introduced by Li et al. and published at AAAI 2023, was shown to outperform prior state-of-the-art models on printed, handwritten, and scene text recognition tasks.

Benchmark performance

Benchmark	Metric	TrOCR-Small	TrOCR-Base (334M)	TrOCR-Large (558M)
IAM (handwritten)	Cased CER	4.22	3.42	2.89
SROIE (printed)	F1	95.86	96.34	96.60

Scene text recognition word accuracy (TrOCR-Base)

Dataset	Accuracy
IIIT5K-3000	93.4
SVT-647	95.2
ICDAR2013-857	98.4
ICDAR2015-1811	86.9
SVTP-645	92.1
CT80-288	90.6

Gigarouter hosts this model as a managed, OpenAI-compatible API. You send image data and receive transcribed text without managing infrastructure or model weights.