Deep face recognition has made significant strides in recent years, largely due to the advancements in convolutional neural networks (CNNs). This article delves into a discriminative feature learning approach that enhances the performance of deep CNNs for face recognition by focusing on minimizing intra-class variations while maximizing inter-class differences. This approach leverages a joint supervision mechanism incorporating a novel center loss in conjunction with the traditional softmax loss.
The CNN architecture used for face recognition experiments. Joint supervision with softmax and center loss is employed.
Enhancing Discriminative Power with Center Loss
The core of this approach lies in the introduction of the center loss. Traditional softmax loss functions primarily focus on separating different classes but often neglect the intra-class compactness. Center loss addresses this by learning a center for each class in feature space and penalizing the distance between deep features and their corresponding class centers. This encourages features from the same class to cluster tightly around their center, leading to more discriminative features.
The implementation details involve a CNN architecture with specific configurations for convolution, local convolution, and fully connected layers. The network is trained using a joint supervision signal combining softmax loss and center loss. The balance between these two losses is controlled by a hyperparameter, allowing for fine-tuning the learning process.
Impact of hyperparameters on face verification accuracy. Performance remains stable across a wide range of values.
Experimental Validation on Benchmark Datasets
The effectiveness of this discriminative feature learning approach is validated through extensive experiments on publicly available face datasets: Labeled Faces in the Wild (LFW), YouTube Faces (YTF), and MegaFace Challenge. The results demonstrate significant improvements in face verification and identification accuracy compared to using softmax loss alone and other existing methods.
Example face image pairs from LFW and YTF datasets. Green frames indicate positive pairs (same person), red frames indicate negative pairs.
Specifically, on the MegaFace Challenge, which tests face recognition performance with millions of distractors, this approach achieves state-of-the-art results under the small training set protocol, significantly outperforming competing methods. Even with a relatively small training dataset, the model exhibits robust performance in challenging scenarios with large-scale distractors. This highlights the power of the center loss in enhancing feature discriminability.
CMC curves comparing different methods on MegaFace with varying distractor sizes.
Conclusion: A Powerful Approach for Deep Face Recognition
The discriminative feature learning approach presented here, utilizing joint supervision with softmax and center loss, offers a compelling solution for enhancing deep face recognition performance. By explicitly addressing intra-class compactness, this approach leads to more robust and discriminative features, enabling superior accuracy in challenging face recognition tasks. This method holds significant promise for real-world applications requiring high accuracy and scalability in face identification and verification.
ROC curves comparing different methods on MegaFace with varying distractor sizes.
Example face images from the MegaFace dataset showcasing the probe set and gallery with distractors.
