Issue Based OCR Error Prediction in Video Streams

This paper increases the reliability of Optical Character Recognition (OCR) systems in natural scene by proposing a novel Image Quality Assessment (IQA) system. We propose to increase reliability based on the principle that OCR accuracy is a function of the quality of the input image. Detected text boxes are analyzed regarding their OCR score and different quality issues, such as blur, light and reflection effects. The novelty of our approach is to model IQA as a classification task, where one class represents high quality elements and each of the other classes represent a specific quality issue. We demonstrate how this methodology allows the training of IQA systems for complex quality metrics, even when no data labeled with the desired metric is available. Furthermore, a single IQA system outputs the quality score as well as the quality issues for a given image. We built on publicly available databases to generate 60k text boxes for each class and obtain 97,1% classification accuracy on a test set of 24k images. We conclude that the learnt quality metric is a valid indicator of common OCR errors by evaluating on the ICDAR 2003 Robust Word Recognition dataset.