Authors: Jacqueline Feild and Erik Learned-Miller
Description: Our word recognition system has three main components. First, we segment images into foreground text and background using a regression based text segmentation technique called bilateral regression. This method models smooth color changes across images and can segment text that varies in color from one part of the image to another. Next, we use a conditional random field (CRF) model with histogram of oriented gradients (HOG) descriptors to find an initial word label based on just appearance and bigram probabilities. Finally, we use an error correction step that incorporates language information from a web-based lexicon of 13.5 million words.