Master Assignment
autoncoders-and-recovering-high-frequency-component
Type: Master EE/CS/HMI
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Description:
Auto encoders(AE) can learn compressed representations of the input data by encoding it to a lower dimensional space and then reconstructing the input again from the encoding. The compressed representation involves the most important attributes of the input data which are enough to reconstruct it as well as possible. According to this definition, this representation can serve as a feature vector for a classification task. The unsupervised learning nature of autoencoders makes them good candidates for cases where it is hard to obtained large amounts of labelled data.
The compression performed by AEs is lossy, and they tend to lose the high frequency components during training. This loss of information can be fatal for some data where details are very important, such as finger veins. Experiments show that AEs lose most of the vein information while learning the low frequency information, such as finger background, almost perfectly. Even though the background information contributes to identity information, it is not enough to achieve a good verification performance on finger veins.
However, the lost high frequency component can be recovered by taking the difference of the input and the reconstruction, and can be re-used in AE training. The goal of this research is to re-utilize this high frequency component for training of a second AE that then can be used to complement the original AE. The work involves exploring AE architectures which utilize the high frequency information, and investigating the relevance of the learned representations on a verification task.