Task specific image text recognition

TLDR: This thesis applies a boosting framework to the character recognition problem, which allows to avoid character segmentation altogether and allows to read blurry, poor quality images that are difficult to segment.

Abstract: This thesis addresses the problem of reading image text, which we define here as a digital image of machine printed text. Images of license plates, signs, and scanned documents fall into this category, whereas images of handwriting do not. Automatically reading image text is a very well researched problem, which falls into the broader category of Optical Character Recognition (OCR). Virtually all work in this domain begins by segmenting characters from the image and proceeds with a classification stage to identify each character. This conventional approach is not best suited for task specific recognition such as reading license plates, scanned documents, or freeway signs, which can often be blurry and poor quality. In this thesis, we apply a boosting framework to the character recognition problem, which allows us to avoid character segmentation altogether. This approach allows us to read blurry, poor quality images that are difficult to segment. When there is a constrained domain, there is generally a large amount of training images available. Our approach benefits from this since it is entirely based on machine learning. We perform experiments on hand labeled datasets of low resolution license plate images and demonstrate highly encouraging results. In addition, we show that if enough domain knowledge is available, we can avoid the arduous task of hand-labeling examples by automatically synthesizing training data