Improving the Character Extraction Rate and Ranking Voice Output in Systems assisting Vision Impaired People Acquire Character Information from the Environment

Abstract

We propose two methods for improving the character extraction accuracy and for ranking the voice outputs in a system that assists vision impaired people acquire character information using a notebook PC and a CCD camera. We identified the region on a sign board using the color of the area from which characters are extracted by local features and found more accurate character strings from projections on the board region. This reduces the number of missed-extractions obtained by local features alone. For ranking purpose, we classified the character strings into headings and content based on the board layout. By comparing the string using a keyword user dictionary, we assigned a high priority to the matched strings. We conducted an experiment for character-string extraction using 164 images (3933 characters). As a result, we obtained a good extraction rate of 91.4%, as compared to only 75.2% in the previous method. For the 121 images, including sign boards with more than one string, we obtained a heading-extraction rate of 55.4% and a correct-ranking rate of 67.7%