COCRID: A Challenging Optical Character Recognition Image Dataset
Robert Grosso, Margaret Pinson · 2023
This memorandum provides technical details for the image quality experiment COCRID: A Challenging Optical Character Recognition Dataset. The design goals of the COCRID dataset are (1) to train no-reference metrics that track the quality of recognized text, (2) to understand characteristics of images that are particularly difficult for Optical Character Recognition (OCR) algorithms, and (3) to develop a metric that responds strongly to the effects of impaired text. The experiment has five environment scenarios and a control to replicate challenging conditions where OCR might be used. This experiment simulates the environment of a mobile scanning application. The experiment photographs source material under a variety of lighting and capture impairments to create a high noise environment. The COCRID contains 984 impaired images and 41 control images. The images are then processed by an OCR algorithm for a result. The resulting string of recognized text is compared with the original to create a character error rate metric. The lessons learned from this dataset will help researchers design datasets for other computer vision algorithms.