Integrating class-dependant tangent vectors into SVMs for handwritten digit recognition

Tangent vectors are one of the best tools for learning variability in handwritten digits. Many research works indicate that tangent vectors provide a significant improvement of accuracy especially when used with SVM classifiers. However, since they are based on the use of affine transformations they substantially extend the runtime. In addition, the user should adequately select the transformations in order to highlight the variability of data. The present work aims to exploit accuracy improvement of tangent vectors while reducing the runtime. Therefore, we investigate the use of tangent vectors that are a priori extracted from training data. The idea is to substitute each pattern by its Tangent Vector Mahalanobis (TVM) distances with respect to all classes. Then, a SVM is trained over TVM values, which contain a priori knowledge and have a smaller size than digit images. Experiments performed on USPS database showed that the proposed approach improves recognition accuracy and allows a huge reduction in the runtime.