Article Preview
Top1. Introduction
In recent years, technologies are developing rapidly. Thus, biological recognition techniques have gained attentions and interests from many researchers. Among those biological recognition techniques, face recognition is a hot researching direction in computer vision and machine learning communities. Face recognition has many characteristics, such as easy to collect data and safe. In addition, face recognition is widely used in e-health area. For example, face recognition is able to accurately and quickly retrieve the medical record of patients when in emergency, which saves the valuable treatment time.
Currently, face recognition approaches can be summarized as four types, which are methods based on geometrical features, methods based on face features, methods based on flexible models, and methods based on Neural Networks (NNs) respectively. Z. Tahira et al. (2018) utilized Principal Component Analysis (PCA) to reduce the dimensions of human faces for implementing the face recognition approach. L. Zhao et al. (2009) employed Support Vector Machine (SVM) to classify human faces and thus solving the face recognition problem. Apart from the methods mentioned above, NNs is a well-known approach being applied to face recognition (Krizhevsky, 2012; Simonyan & Zisserman, 2014). It has many advantages, such as strong ability of non-linear regression, robust and good performance in effectiveness and efficiency.
This paper proposed a novel approach employing Convolutional Neural Network (CNN) to extract features from human faces. Next, the extracted features are used as training data to train the CNN model. Last, the trained model is utilized to predict the final result. Our contributions are concluded as follows:
- 1.
Through separating convolutional kernels, the parameters in the CNN are reduced, and thus the depth of the network model is increased, which leads to a lower execution time and higher accuracy;
- 2.
Utilizing Batch Normalization to reduce the training time;
- 3.
Adopting ReLU and Dropout to reduce over-fitting.
Top2. Convolutional Neural Network
Convolutional Neural Network (CNN) has the characteristics of sparse connection and parameters sharing. These two features help reducing a large number of parameters, and thus accelerating the learning rate. Therefore, CNN can be applied into the face recognition problem effectively since it reduces the training time and increases the accuracy of the CNN model.
In terms of selecting and tuning the convolutional kernel and activation function, there are several features should be considered: