Text detection and classification in natural images is important for many computer vision
applications. Here, to detect text, we find connected
components based on consistency in stroke width, chain
them based on their relative spatial positions, and use a
text classification engine to filter chains with low classi!cation con!dence scores. The algorithm performs
well using the ICDAR 2015 competition images, with
72.4% precision and 71.0% recall, but struggles under
certain imaging conditions or types of text. Finally, we
discuss ways to improve the algorithm to yield better
performance.