Image Text Detection Using a Bandlet-Based Edge Detector and Stroke Width Transform

A slew of semantic image content analysis techniques are specialized in extracting text embedded in images since it is a vital source of semantic information. A robust text detection step is the basic requirement for a scheme designed to extract text information from images. Text detection is still a challenging issue due to unconstrained color, sizes, alignments of characters, lighting and also various shapes of fonts, even though various methods have been proposed in the past years [2]. Existing text detectors can be broadly classified into two main groups: texture (also called region) based and connected component (CC) based methods. The general scheme of our proposed method consists in producing (Fig. 1) the image edge map and then finding CCs based on stroke width transform (SWT) [1] guided by the generated edge map. Next, precise feature vectors are formed using the properties of CCs from SWT and pixel domain. An unsupervised clustering is performed on the image CCs to detect the candidate text CCs. Finally, text candidate components are linked to form text-words. The method is considered as a CC-based technique and the contribution is twofold: 1) Since accurate edge maps drastically enhance SWT results, a precise edge detection approach adaptive to text-regions is proposed by employing the bandlet transform. 2) A feature vector based on text properties and stroke width values is employed in k-means clustering in order to detect text CCs. Bandlet transform [3] effectively represents the geometry of an image. The image coefficients are dyadically segmented in squares S for polynomial flow approximation of the geometry before the bandletization process. Since the image coefficients are all warped along local dominant flows in the bandlet transform, the final bandlet coefficients generated for each segmentation square S have the form of approximation, and highpass filtering values appear in the wavelet transform of a 1D signal. We benefit from the bandlet-based resulting 1D high-pass frequency coefficients that are adapted to the directionality of the edge that exists in each segmentation square S in order to find a binary map of the edge positions in the image. Since the approximation part of the bandlet transform resulting coefficients consists of coarse information of the original signal, we discard it and only process the high-pass coefficients. The first-order derivatives of the fine-detail bandlet coefficients are computed. Then, local maxima of the resulting gradient signal are found using a contextual filter as follows: