Hyper-parameter Determination of CNN Classifier for Head Pose Estimation of Three Dimensional Degraded Face Images

This paper presents the evaluation of parameters for head pose estimation using Convolutional Neural Network (CNN) towards the degraded images. Head pose estimation is one of the important factor for three dimensional face recognition system. Due to its superiority, Convolutional Neural Network (CNN) has been used as a head pose estimator, however, its performance is significantly dropped when the input face images is exposed to noises. As the CNN comes with different choices of pooling layer, two different experimental setups are created with similar architecture and training condition but using a different type of pooling layer. After learning, the CNN are tested with another five different testing datasets to monitor the effects of various particular noises, such as: Gaussian noise, Salt-Pepper, and Speckle. Result of the experiments shows that the usage of max pooling significantly lowering the performance of the CNN, compared to the system with average pooling layer.