Detecting Texts of Arbitrary Orientations in 1 Natural Images

5 Texts in a natural image directly carry rich high-level semantic information about a scene, which can 6 be used to assist a wide variety of applications, such as image understanding, image indexing and search, 7 geolocation or navigation, and human computer interaction. However, most existing text detection and 8 recognition systems are designed for horizontal or near-horizontal texts. With the increasingly popular 9 computing-on-the-go devices, detecting texts of arbitrary orientations from images taken by such devices 10 under less controlled conditions has become an increasingly important and yet challenging task. In this 11 paper, we propose a new algorithm to detect texts of arbitrary orientations in natural images. Our algorithm 12 is based on a two-level classification scheme and utilize two sets of features specially designed for 13 capturing both intrinsic and orientation-invariant characteristics of texts. To better evaluate the proposed 14 method and compare it with other existing algorithms, we generate a more extensive and challenging 15 dataset, which includes various types of texts in diverse real-world scenes. We also propose a new 16 evaluation protocol, which is more suitable for benchmarking algorithms designed for texts of varying 17 orientations. Experiments on conventional benchmarks and the new dataset demonstrate that our system 18 compares favorably with the state-of-the-art algorithms when handling horizontal texts and achieves 19 significantly enhanced performance on texts of arbitrary orientations in complex natural scenes. 20