| Summary: | Face recognition has attracted particular interest in biometric recognition with wide applications in security, entertainment, health, marketing.
Recent years have witnessed rapid development of face recognition technique in both academic and industrial fields with the advent of (a) large amounts of available annotated training datasets, (b) Convolutional Neural Network (CNN) based deep structures, (c) affordable, powerful computation resources and (d) advanced loss functions. Despite the significant improvement and success, there are still challenges remaining to be tackled.
This thesis contributes towards in the wild face recognition from three perspectives including network design, model compression, and model explanation.
Firstly, although the facial landmarks capture pose, expression and shape information, they are only used as the pre-processing step in the current face recognition pipeline without considering their potential in improving model's representation. Thus, we propose the ``FAN-Face'' framework which gradually integrates features from different layers of a facial landmark localization network into different layers of the recognition network. This operation has broken the align-cropped data pre-possessing routine but achieved simple orthogonal improvement to deep face recognition. We attribute this success to the coarse to fine shape-related information stored in the alignment network helping to establish correspondence for face matching.
Secondly, motivated by the success of knowledge distillation in model compression in the object classification task, we have examined current knowledge distillation methods on training lightweight face recognition models. By taking into account the classification problem at hand, we advocate a direct feature matching approach by letting the pre-trained classifier in teacher validate the feature representation from the student network. In addition, as the teacher network trained on labeled dataset alone is capable of capturing rich relational information among labels both in class space and feature space, we make first attempts to use unlabeled data to further enhance the model's performance under the knowledge distillation framework.
Finally, to increase the interpretability of the ``black box'' deep face recognition model, we have developed a new structure with dynamic convolution which is able to provide clustering of the faces in terms of facial attributes. In particular, we propose to cluster the routing weights of dynamic convolution experts to learn facial attributes in an unsupervised manner without forfeiting face recognition accuracy. Besides, we also introduce group convolution into dynamic convolution to increase the expert granularity. We further confirm that the routing vector benefits the feature-based face reconstruction via the deep inversion technique.
|
|---|