Real time image enhancement and segmentation for sign/text detection

We describe a pc−based low cost visual system that can detect and extract text regions in visual signs in the scene and recognize them for location awareness. It employs a multi resolution image enhancement and segmentation methods based on symmetric neighborhood filter and hierarchical connected component analysis to extract written information on signboards which appears in the scene. The multi resolution approach allows us to remove the undesired segments by analyzing their features, e.g. size, contrast, shape on low−resolution, remove the perspective distortion and correct the orientation for better detection using signboard boundary lines and corners on medium−resolution, and extract the words, characters on high−resolution images for better recognition. As the signboard may not contain only text but also may contain non−text illustrations, the systems classifies each text region inside the signboard as text−only and sign−only using SVM based classifiers, where shape histograms, a distribution of relative positions of the pixels, for each text region is used as a feature vector. Experiment results show the robustness of our region−based detection algorithm over pixel−based algorithms.