Monotonic Linear Spline Activation Functions And Face Recognition With Rotation Angle And Image Quality Score

Recent years,neural networks have achieved great breakthrough in many fields.Activation function plays a significant role in neural networks,ReLU is essential for re-cent state-of-the-art deep neural networks.Face recognition achieves exceptional success thanks to the emergence of deep neural networks.However,many contemporary face recognition models still perform relatively poor in processing profile faces,and the quality of each sample is not exactly same,samples with poor quality will hurt the metric.In this study,we want to propose a high quality activation function and face recognition model.In this work,we will study image classification from two aspects.First,we propose Binary PReLU and Leaky PReLU activation function that generalizes the parametric rectified unit.Binary PReLU and Leaky PReLU improve model fitting with little compu-tational cost.Second,we improve the pose robust and with quality score face recognition model.Exploit the head rotation and middle feature,we can get the final feature output.We combine the loss and predicted quality score to reduce the requirement of dataset and obtain a better loss function.The first part of the thesis mainly introduces the background and status of the activation function and face recognition algorithm,and the significance and content of the research topic.In the second part,we will present the Binary PReLU and Leaky PReLU detailedly.For the third part of this paper,we join the head rotation and quality score into our model In the last part of this paper,we conduct many experiments on common datasets and nerworks,then show the experimental results.