When Face Recognition Meets with Deep Learning: an Evaluation of Convolutional Neural Networks for Face Recognition (1504.02351v1)

Abstract: Deep learning, in particular Convolutional Neural Network (CNN), has achieved promising results in face recognition recently. However, it remains an open question: why CNNs work well and how to design a 'good' architecture. The existing works tend to focus on reporting CNN architectures that work well for face recognition rather than investigate the reason. In this work, we conduct an extensive evaluation of CNN-based face recognition systems (CNN-FRS) on a common ground to make our work easily reproducible. Specifically, we use public database LFW (Labeled Faces in the Wild) to train CNNs, unlike most existing CNNs trained on private databases. We propose three CNN architectures which are the first reported architectures trained using LFW data. This paper quantitatively compares the architectures of CNNs and evaluate the effect of different implementation choices. We identify several useful properties of CNN-FRS. For instance, the dimensionality of the learned features can be significantly reduced without adverse effect on face recognition accuracy. In addition, traditional metric learning method exploiting CNN-learned features is evaluated. Experiments show two crucial factors to good CNN-FRS performance are the fusion of multiple CNNs and metric learning. To make our work reproducible, source code and models will be made publicly available.

• The paper introduces three CNN architectures for face recognition and finds that CNN-M strikes an optimal balance between network complexity and training data size.
• It shows that reducing the dimensionality of CNN-learned features maintains recognition accuracy while enhancing storage efficiency and computation speed.
• The study validates the effectiveness of network fusion combined with Joint Bayesian metric learning, significantly boosting face recognition performance.

Evaluation of Convolutional Neural Networks for Face Recognition

The paper "When Face Recognition Meets with Deep Learning: an Evaluation of Convolutional Neural Networks for Face Recognition" presents a comprehensive paper on the application and efficacy of Convolutional Neural Networks (CNNs) in the domain of face recognition. Recognizing the successes of CNNs in recent years, the authors aim to evaluate different CNN architectures specifically designed for face recognition tasks, and address the often neglected question of why and how certain CNN architectures excel in this domain.

Summary of Key Contributions

1. Design and Evaluation of CNN Architectures: The authors introduce three CNN architectures (CNN-S, CNN-M, and CNN-L) of varying complexities and conduct experiments using the Labeled Faces in the Wild (LFW) dataset. The CNN-M architecture emerges as the most effective, outperforming both CNN-S and CNN-L, likely due to an optimal balance between network complexity and training data size.
2. Analysis of Feature Properties: The paper provides an insightful evaluation of the learned features’ properties, demonstrating that the dimensionality of CNN-learned face features can be significantly reduced without compromising recognition accuracy. This implies feasible storage reductions and computational efficiency, advantageous for large-scale applications.
3. Effectiveness of Network Fusion and Metric Learning: The paper substantiates the benefits of network fusion, where multiple CNNs trained on different facial regions and scales improve the recognition performance. It is established that this enhanced feature representation significantly outperforms single-network models. Additionally, the application of metric learning, specifically the Joint Bayesian (JB) method, further optimizes the classification accuracy.
4. Comprehensive Evaluation of Implementation Choices: Through extensive experimentation, the authors analyze various implementation choices such as input image type (gray-scale vs. color), data augmentation techniques, and distance metrics for feature comparison. The cosine distance is highlighted as the superior metric for evaluating CNN-learned features among the measures tested.

Implications and Future Directions

While the research offers valuable insights into architectural design factors and their relative impacts on face recognition, it also prompts several questions requiring further exploration. Notably, the promising results from network fusion and metric learning techniques suggest future work could focus on optimizing and expanding these methodologies. The understanding that certain combinations of network designs and implementation strategies yield superior performance can guide the creation of more sophisticated and efficient face recognition systems.

Going forward, this research potentially lays a foundation for further theoretical analysis to complement empirical findings. This would enrich the understanding of convolutional architectures not just for face recognition but for broader pattern recognition tasks, fostering advancements in adaptive CNN architecture design.

Moreover, with the source code and models made publicly available, the authors provide a substantial resource for reproducibility and future innovation in the field. This openness supports experimentation and potential enhancements by the research community, furthering the development of robust face recognition solutions.

In summary, this paper represents a methodical evaluation of CNN methodologies, improving our understanding of design and implementation choices in face recognition systems. It challenges researchers to further refine these networks and continues to question the theoretical underpinnings of why these systems operate optimally, thus guiding future efforts in AI and computer vision landscapes.