Optical character recognition based document image quality assessment

Optical Character Recognition (OCR) systems play a crucial role in digitizing documents. However, their performance significantly deteriorates when handling low-quality images. Even advanced OCR systems struggle if the input is visually or structurally poor. Therefore, achieving high OCR accuracy requires assessing document image quality in terms of how well characters can be recognized, not just visual clarity. In this work, we propose a Document Image Quality Assessment (DIQA) model that predicts OCR accuracy without requiring the actual execution of an OCR engine. To assess document image quality for OCR performance, twelve distinct features are extracted that capture various aspects of sharpness, focus, edge clarity, and structural distortion. Instead of relying on subjective human opinion scores, we generate labels by measuring the actual OCR accuracy using modern engines like PaddleOCR and Keras OCR. These accuracy scores, calculated using the Levenshtein distance, serve as ground truth labels for training. Using the extracted features and corresponding OCR-based labels, we train the machine learning models to learn the relationship between image characteristics and OCR performance. The proposed models are evaluated using statistical metrics such as RMSE, PLCC, and SROCC to determine the most effective predictor. Our experiments demonstrate the importance of using OCR scores as labels, and the results show that our approach yields improved performance compared to existing baseline methodologies.