Design and Evaluation of Features That Best Define Text in Complex Scene Images

In this paper we explore features for text detection within images of scenes containing other background objects using a Support Vector Machine (SVM) algorithm. In our approach, the Haar-like features are designed and utilised on banks of bandpass filters and phase congruency edge maps. The designed features with SVM leverages the properties of text geometry and colour for better differentiation of text from its background in real world scenes. We also evaluate the contributions of the features to text detection by the SVM coefficients, which leads to time-efficient detection by using an optimal subset of features.