ADAPTIVE NEURAL NETWORK CLASSIFIER FOR EXTRACTED INVARIANTS OF HANDWRITTEN DIGITS

Abstract

We propose an adaptive activation function of neural network classifier for isolated handwritten digits that undergo basic transformations. The utilized network is a backpropagation network with sigmoid and arctangent activation functions. The performance of network with both activation functions is compared. The results show that the network applying an adaptive activation function between layers converged much faster compared to non-adaptive activation functions with 50% iterations reduction. In this study, we also present experimental results of feature extraction between Zernike and d-geometric for better feature representations. Results show that Zernike features are better at representing isolated handwritten digits compared to d-geometric features with an accuracy up to 87%.