Text and Symbol Extraction in Traffic Panel from Natural Scene Images

Abstract

Traffic panels contain rich text and symbolic information for transportation and scene understanding. In order to understand the information in panels, fast and robust extraction of the text and symbol is a crucial and essential step. This problem cannot be solved using generic scene text detection methods due to the special layout characteristics, especially in Chinese panels. In this paper, we propose a fast and robust approach for Chinese text and symbol extraction in traffic panels from natural scene images. Given a traffic panel in natural scene, Contrasting Extremal Region (CER) algorithm is applied to extract character candidates which are further filtered by boosting classifier using Histogram Orientation Gradient Features. Since Chinese characters often consist of multiple isolated strokes, a hierarchical clustering process of stroke components is carried out to group isolated strokes into characters using the detected characters as seeds. Next, the Chinese text lines are formed by Distance Metric Learning (DSL) method. In consideration that traffic symbols do not possibly appear in the location of texts, symbols are extracted using two stages boosting classifier after text detection. Experimental results on real traffic images from Baidu Street View demonstrate the effectiveness of the proposed method.