# DTrOCR: Decoder-only Transformer for Optical Character Recognition

**Authors:** Masato Fujitake  
**arXiv:** [2308.15996](https://arxiv.org/abs/2308.15996)  
**Date:** 2023-08-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.15996v1/

## Key points

- Traditional text recognition methods use an encoder-decoder structure
- Masato Fujitake introduced the Decoder-only Transformer for Optical Character Recognition (DTrOCR)
- DTrOCR utilizes a decoder-only Transformer model pre-trained on a generative language model
- DTrOCR departs from the conventional encoder-decoder framework and focuses on decoder-only architecture
- DTrOCR outperforms existing state-of-the-art methods in recognizing printed, handwritten, and scene text in English and Chinese languages

## Summary

, , , , 

In the field of text recognition, traditional methods have relied on an encoder-decoder structure where the encoder extracts features from an image and the decoder generates recognized text based on these features. However, a recent study by Masato Fujitake introduces a novel approach called the Decoder-only Transformer for Optical Character Recognition (DTrOCR). This method utilizes a decoder-only Transformer model that leverages a generative language model pre-trained on a vast corpus. The key innovation of DTrOCR lies in its departure from the conventional encoder-decoder framework, opting instead for a decoder-only architecture. By harnessing the power of a generative language model originally designed for natural language processing tasks, DTrOCR aims to explore the effectiveness of such models in the realm of computer vision and text recognition. Through comprehensive experiments and evaluations, Fujitake demonstrates that DTrOCR significantly outperforms existing state-of-the-art methods across various types of text, including printed, handwritten, and scene text in both English and Chinese languages. The results highlight the superior performance and efficiency of DTrOCR in accurately recognizing text from images compared to traditional approaches. Furthermore, this groundbreaking research has been recognized with acceptance at WACV2024, underscoring its significance and potential impact on advancing optical character recognition technologies. Overall, Fujitake's work showcases how innovative techniques like DTrOCR can push the boundaries of text recognition capabilities and pave the way for more efficient and accurate solutions in this domain.