A robust algorithm for text region detection in natural scene images

In this paper, a new method for detecting text regions in natural scene images is presented. The proposed algorithm is based on the segmentation of objects in a scene, followed by the identification of text objects by a support vector machine (SVM). First, to segment objects in the scene, the input image is separated into chromatic and achromatic regions according to the distribution of red, green and blue (RGB) elements, and different clustering algorithms are applied. Second, each object is transformed into the wavelet domain for multi-resolution analysis, and moment features of the wavelet coefficients are used in the SVM for the classification of text objects. The proposed approach provides robustness to non-uniform illumination by using different clustering algorithms according to the characteristics of the colour components in the segmentation. Also, moment features, used for classification, are invariant to the size, direction, shape and other properties of texts. Experimental results demonstrate the effectiveness of this approach.

I Introduction

In natural scene images, texts provide highly useful information and can provide a key for understanding the image content. To extract semantics from texts in a scene, text regions should be identified first. Text region detection can be useful in many applications, such as roadsign identification and the interpretation of documents such as geographic maps or engineering drawings. There has been extensive research on text detection and recognition in recent years. However, because of the extreme diversity of text patterns together with illumination variations, distortions and other noise, text detection within a natural scene remains a difficult and challenging task. Many approaches have been proposed for the segmentation, detection and recognition of texts from document images, video sequences or natural scene images.