Characterizing and Correcting Capture-Specific Defects in Images Taken by Visually Impaired Users to Improve OCR-Based Text Recognition by Comparing OCR Engines.

OCR, optical character recognition, has proven to be effective in reading text from images. However it can be a non-trivial task, especially from the point of view of visually impaired users, due to possible defects in the image. We explore whether OCR preprocessing should be tailored to the actual defect present in an image or whether it is always optimal to apply the same preprocessing, regardless of image defects. We analyze 2000 images from the VizWiz corpus, each with four possible types of defects: blur, tilt, low illumination, and overexposure. We compare three conditions: the raw image, a standard preprocessing, and a defect-specific preprocessing. We evaluate these using two OCR engines, Tesseract and EasyOCR. The error rates in terms of CER and WER are computed on a quality-screened set of 177 images with reference text. With different results for each OCR engine, the Tesseract’s best CER and WER were achieved with the defect-specific preprocessing, whereas for EasyOCR the best results were obtained with the standard preprocessing. Therefore, it is not possible to identify a single best preprocessing for all images and OCR engines.