OCR-Diff: A Two-Stage Deep Learning Framework for Optical Character Recognition Using Diffusion Model in Industrial Internet of Things
Chaewon Park, Vikas Palakonda, Sangseok Yun, Il‐Min Kim, Jae‐Mo Kang
Abstract
Optical character recognition (OCR) is one of the key enabling technologies in industrial internet-of-things (IIoT) for extracting and utilizing useful textual information, but it is technically challenging due to poor environmental conditions. To deal with such challenges, in this letter, we propose a novel two-stage deep learning framework for OCR using a generative diffusion model, namely, OCR-Diff. In the first stage, our customized conditional U-Net is pre-trained jointly with a feature extractor with the aid of the forward diffusion process such that the quality of a low-resolution text image is improved via the reverse diffusion process. In the next stage, the pre-trained conditional U-Net and feature extractor are jointly fine-tuned for an off-the-shelf text recognizer to precisely recognize the texts in the image. Experimental results on TextZoom datasets substantiate the superiority and effectiveness of the proposed scheme.