Hi,
In the texture lecture, we talk about using the discrete Fourier transform (DFT) to classify different textures. It is not clear to me what type of classifier is used in practice. The DFT returning a high dimensional signal. It is fairly straightforward to use a CNN for that task but how would "old-style" classifier deal with it ? Is the map simply flattened into a very long feature vector ?
Thanks in advance.
Hi,
Yes, simply flattening into a very long feature vector is an option. But, back in the days, feature engineering played a big part in classifier design. An engineer would try to find which frequencies contribute most to making an accurate classification. Then, only those frequency ranges will be used in the feature vector (instead of the full spectrum). Along with these DFT features, he might decide to use other features and will concatenate them with this feature vector. Finally, it would be fed into a standard classifier like SVM.
Follow the pulses of EPFL on social networks
© 2023 EPFL, all rights reserved