Detection and Recognition for Reading Text in Images

Abstract

Detection And Recognition for Reading Text in Images is a difficult but important problem. It can be summarized as: how to enable a computer to recognize letters and digits from a predefined alphabet, possibly using contextual information. Various attempts at solving this problem, using different selections of features and classifiers, have been made. Human performance has been achieved in accuracy by automated text recognition systems, and has been bypassed in speed for the case of single size, single font, high quality, known layout, known background, text. When one or more of the above parameters are changed, the problem becomes increasingly difficult. In particular, attaining human performance in recognizing cursive script of varying size, varying style, unknown layout, unknown background is far from the reach of todays' algorithms, despite the continuous research effort for almost four decades. In this report, we analyze the problem in detail, present the associated difficulties, and propose a coherent framework for addressing automated text recognition. A lot of people like to say that the world is overwhelmed with information that is still harder and harder to deal with, both for individual humans living in the overwhelmed world and for the technology they use. Popularity of mobile devices equipped with cameras has influenced peoples' lives in many ways recently. One of these changes is that people started to take photos as notes about things which are not of visual nature as opening hours or traffic schedules. Taking a picture of the signs became a very convenient way of storing such information, however later retrieval of such "photographic notes" with any meta-data may became very time consuming.