Evaluation of face recognition algorithms under noise

Almatarneh, Ansam (2019) Evaluation of face recognition algorithms under noise. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (2MB)

Abstract

One of the major applications of computer vision and image processing is face recognition, where a computerized algorithm automatically identifies a person’s face from a large image dataset or even from a live video. This thesis addresses facial recognition, a topic that has been widely studied due to its importance in many applications in both civilian and military domains. The application of face recognition systems has expanded from security purposes to social networking sites, managing fraud, and improving user experience. Numerous algorithms have been designed to perform face recognition with good accuracy. This problem is challenging due to the dynamic nature of the human face and the different poses that it can take. Regardless of the algorithm, facial recognition accuracy can be heavily affected by the presence of noise. This thesis presents a comparison of traditional and deep learning face recognition algorithms under the presence of noise. For this purpose, Gaussian and salt-andpepper noises are applied to the face images drawn from the ORL Dataset. The image recognition is performed using each of the following eight algorithms: principal component analysis (PCA), two-dimensional PCA (2D-PCA), linear discriminant analysis (LDA), independent component analysis (ICA), discrete cosine transform (DCT), support vector machine (SVM), convolution neural network (CNN) and Alex Net. The ORL dataset was used in the experiments to calculate the evaluation accuracy for each of the investigated algorithms. Each algorithm is evaluated with two experiments; in the first experiment only one image per person is used for training, whereas in the second experiment, five images per person are used for training. The investigated traditional algorithms are implemented with MATLAB and the deep learning algorithms approaches are implemented with Python. The results show that the best performance was obtained using the DCT algorithm with 92% dominant eigenvalues and 95.25 % accuracy, whereas for deep learning, the best performance was using a CNN with accuracy of 97.95%, which makes it the best choice under noisy conditions.

Actions (login required)

View Item