A robust modular wavelet network based symbol classifier

This paper presents a robust automatic shape classifier using modular wavelet networks (MWNs). A shape descriptor is constructed based on a combination of global geometric features (modified Zernike moments and circularity features) and local intensity features (ordered histogram of image gradient orientations). The proposed method utilizes a supervised modular wavelet network to perform shape classification based on the extracted shape descriptors. Affine invariance is achieved using a novel eigen-based normalization approach. Furthermore, appropriate shape features are selected based on the inter- and intra-class separation indices. Therefore, the proposed classifier is robust to scale, translation, rotation and noise. Modularity is introduced to the wavelet network to decompose the complex classifier into an ensemble of simple classifiers. Wavelet network parameters are learned using an extended Kalman filter (EKF). The classification performance of proposed approaches is tested on a variety of standard symbol data sets (i.e., mechanical tools, trademark, and Oriya numerals) and the average classification accuracy is found to be 98.1% which is higher compared to other shape classifier techniques.