Versatile Text Extraction System for Text-to-Speech Reading Assistant Camera

Wearable camera device translating the text in the scene into speech is one of the most anticipated devices for the visually-impaired. The users would probably want to read any text using such a device. Although various scene text extraction methods have been developed so far, the target objects are most often limited to simple signboards, small memos, etc. We propose a versatile scene text extraction method that can handle a wide variety of targets including complex signboards with many text lines. Experimental results show that our system runs at a video rate and can extract densely arranged text lines even with some distortion and shading. A locally-adaptive binarization technique contributes to the better quality of extracted text images.