Text segmentation in ancient topographic maps and floor plans with support vector data description

Images of ancient maps and floor plans can present a great challenge for common character recognition tools. Besides the damage caused by time and handling, these documents have an important part of their information described graphically. In most examples, drawings of rivers or walls occupy most part of the document. Usually, text has different styles, sizes and orientations with possible overlapping with graphics. This paper presents a new method for text segmentation in images of ancient topographic maps and floor plans that uses a machine learning algorithm specialized in novelty detection to decide which components of the image are textual. Despite using artificial text examples for training, the method is able to outperform other state-of-the-art methods when applied to real images.