UC San Diego

Text is a fundamental medium for visual communication. Methods to automatically read text, known as Optical Character Recognition (OCR), have a long history and by the 1960's found significant commercial application. Though in the past the focus of OCR technology has been in the domain of scanned books and documents, there are significant new sources of images resulting from the widespread use of digital cameras, mobile phones, vehicle- mounted cameras, wear- able cameras, and more. This new flood of images continues to grow at a rapid pace and presents the OCR problem with new challenges and opportunities. In contrast to the well-controlled scanned document environments of the past, these new data sources present a fundamental technical challenge to existing OCR engines due to the nature of being acquired in unconstrained environments. These factors present us with a unique situation : OCR is historically a desirable technology, more images are being collected now than ever before, and due to the nature of this new wave of data, conventional techniques to solve the problem are unfit. In a parallel literature, generic object recognition research has emerged to deal with the challenge of visual recognition in complex real world environments. While those advances pushed the field's ability to recognize a wide range of objects, very little had been done to translate this paradigm to recognition of one of the most important visual objects : text. In this dissertation, we aim to bridge the gap between the new learning-based techniques of object recognition to the problem of recognizing text in the wild. While the expectation of non -practitioners may have been that text recognition was a solved problem (due to its past success), we both break that perception and present a path for future progress in this challenging problem of considerable practical interest