A Low Error Face Recognition System Based on A New Arrangement of Convolutional Neural Network and Data Augmentation

This paper represents a low error face recognition system constructed on convolutional neural network structure. The proposed method introduces a new layer arrangement for the CNN with added normalization layers. In addition, a data augmentation step consisting of vertical flip, scaling, rotation, and shift is implemented into the framework of our system to improve the accuracy. This augmentation helps the system to overcome the issue of having low number of samples per individuals in our dataset. The support vector machine (SVM) and Softmax are considered as classifiers of the proposed system, and results are evaluated for both classification methods. The system is tested on the ORL face image dataset. In our experiment SVM showed a higher accuracy compared to Softmax. The results compared to other existing face recognition methods show better performance and higher accuracy of the proposed system. The proposed method is evaluated with %50 of the dataset as training samples and the rest as test samples by random. Our technique achieves %98.64 with Softmax and %99.7 with SVM recognition rate.