Extraction of Characters on Signboards in Natural Scene Images by Stump Classifiers

We present a method to detect characters on signboards in natural scene images. For many applications, both classifier with small computational cost and the efficient feature set, which gives rise to accurate recognition are required. Texture based features are often used for target detection. It has been also shown that the shape of the intensity distribution is often useful for character extraction. The intensity distribution in the character regions is often different from the unimodal distribution. We measure the discrepancy between the observed region and the normal distribution by skewness and kurtosis. We use these statistics along with the texture based features. Character regions in a natural scene image are detected by using the linear combination of stump classifiers, each of which sees only one component of multidimensional feature vector. Selection of a feature component for each stump and determination of coefficients of linear combination are carried out by AdaBoost. We experimentally show the effectiveness of the proposed method.