Gradient normalization based gradient feature extraction

A gradient normalization based gradient feature is extracted for handwritten character recognition. First, the gradient of original character image is calculated; then the gradient is normalized making use of piecewise linear normalization function; last, a gradient feature vector is constructed from the normalized gradients. The overall framework is novel since it is based on gradient normalization; then, the piecewise linear normalization function is also new which is effective for rectifying deformations and has merit of computation simplicity. They enhance the robustness of gradient feature to handwritten character deformations and promising empirical results are achieved by our proposed schemes.