Navigation auf uzh.ch
Navigation auf uzh.ch
The task of face recognition is to compare two facial images and assess if they are from the same or from different subjects. In the era of deep learning, face recognition relies on the comparison of deep features that are extracted from a face via a pre-trained deep neural network such as ArcFace Usually, this extraction process highly relies on face alignment, i.e., facial landmarks have to be detected in order to put the face upright and at a specific location inside of the image.
The goal of this Bachelor's thesis is to assess which kinds of perturbations of image alignments and/or blur can be used in order to still obtain good performance. A similar study has been performed using traditional face recognition algorithms in Impact of Eye Detection Error on Face Recognition Performance. For example, different rotations, scales and shifts of the face can be applied to the image that is presented to the network. From the deep features extracted from these perturbations, the one that is working best can be determined by an algorithm that is to be designed in this Bachelor's thesis. One such criterion might be the Euclidean length of the deep feature representation -- better representations usually have higher Euclidean lengths.
By training our AFFFE network to be more robust to face misalignment (using the technique from AFFACT), we removed the dependence on this landmark detection step and we could directly make use of the output of the face detector in order to crop the face. As we have shown in AFFACT, it can be beneficial to use test data augmentation, i.e., to apply different perturbations during the face alignment step in order to improve the task (which was facial attribute classification in that case). Additionally, deep features extracted from many of these transformations can be joined into a single representation of the face, for example, by averaging. Tasks of this Bachelor's thesis would include to determine a certain useful subset of transformations, and to maybe find a better strategy than the simple average.
Experiments will be conducted on the Labeled Faces in the Wild (LFW) dataset following its pre-defined evaluation protocol. It is recommended to make use of the Biometric Recognition packages, which are part of the Bob toolbox and which also include means of image alignment.
This thesis can be extended to a Master's thesis by training a better version of the AFFFE network on a larger dataset using the technique presented in AFFACT, which has recently been re-implemented as a Master's project in my lab.
