An Optical Character Recognition Technique for Extracting Text from Blurred and Low-Quality Images Using TrOCR

Visually impaired individuals struggle with blurry and low-quality images, and it is tough for senior citizens to learn by reading printed materials. The research is intended to help older people find the text they want even with low-quality pictures or vision impairment. The principal target of that research is extending the capabilities of existing OCR technologies. This research demonstrates a novel method for extracting text from blurred and low-quality photos, which is a serious difficulty for elderly users, using Transformer-based Optical Character Recognition (TrOCR). For this text extraction, we use the VisionEncoderDecoder model framework with weight initialized from TrOCR in the transformers library. We took two dataset types, the quote images dataset and IIIT 5K benchmark dataset. The focus is on evaluating the performance and accuracy of the proposed methodology in diverse contexts, quote text images, and obscured text in scene images. The performance of the proposed methodology is evaluated on benchmark datasets IIIT 5K-word dataset. The proposed methodology, utilizing TrOCR with Sequence to Sequence (Seq2Seq) architecture achieves high accuracy rates of 96.8% on IIIT 5K, 83% on Quote Text Image dataset indicating its robustness in handling complex text recognition tasks. This study highlights the strengths and limitations of current OCR methodologies, providing insights into the effectiveness of the proposed methodology in the real world applications.