Face Recognition Using Robust Convolutional Neural Network

Face recognition is still a challenging issue especially when the images contain various kinds of occlusions, illumination variations, and poses. We propose robust Convolutional Neural Network (CNN) with the new cost function including the back propagated error and gradient of the hidden neuron penalty. The gradient penalty follows Hebb’s learning rule multiplied by the derivative of sigmoid function, which avoids the weights from drastically changing when it feedbacks the small variations of the output error to the input layer. The proposed method compared with I2DKPCA and conventional CNN shows that the proposed approach outperforms existing state of art methods.