I am trying to use wavelet transform as part of pooling operation in Convolutional Neural Network for Facial emotion classification.As part of this I tried with simple Haar wavelet but it doesn't give me good results not even better than max pooling operation. So I wanted to know which wavelet transform would be better to use compared to Haar or is it really worth using wavelet transform for pooling operation? I am quite confused with this and also it is the first time I am trying to explore more on wavelets and use them in my project in order to get better accuracy. Please suggest me on this or please direct me whether I am going in right direction or not.
Wavelet transform is not a good choice for facial recognition. The reason is a WT can model textures only in a limited number of directions. For example Haar wavelet can do so only in 3 directions : horizontal, vertical, diagonal, which is not sufficient for modeling complex facial features. The most promising features for facial features have been found to be derived from eigenfaces (eigen decomposition of facial images). For more info explore the links below :
https://en.wikipedia.org/wiki/Eigenface
http://www.face-rec.org/algorithms/pca/jcn.pdf
http://www.face-rec.org/interesting-papers/
https://onionesquereality.wordpress.com/2009/02/11/face-recognition-using-eigenfaces-and-distance-classifiers-a-tutorial/
https://jeremykun.com/2011/07/27/eigenfaces/
Chapter Facial Expression Recognition
https://www.hindawi.com/journals/mpe/2019/9185481/
https://in.mathworks.com/matlabcentral/fileexchange/33325-eigenface-based-facial-expression-classification
Chapter Facial Expression Recognition Using Image Processing Techniq...
https://sites.cs.ucsb.edu/~mturk/Papers/mturk-CVPR91.pdf
Ranjan Parekh Thank you so much let me explore on the stuff that you have provided.
Ranjan Parekh there is one thing which I would like to ask may be my question would be wrong but still I would like to clear my doubt. What I read in the articles related to wavelet, People are using approximation coefficients that we get in LL (low pass band using low pass filters) and ignoring others. Because in this band wavelets capture and preserve the details of the image. So I think that for facial emotions also if we consider LL band then it should work? Please correct me if I am wrong.
The approximation coefficient is only a low-resolution version of the original image. Apart from noise reduction to some extent it will not provide any extra feature for classification. The detail coefficients on the other hand will provide information regarding texture orientations along specific directions, which can be used for classification, if certain facial emotions generate strong textures along specific directions. Note however that all coefficients are 2D matrices, so some statistical measures would need to calculated from each and formulated as a feature vector before it can be fed to a classifier.
Asma Khatun Thanks I will try for it. I am actually trying to use wavelets in pooling operation of CNN like instead of maxpooling I can use wavelet pooling where I can use both low pass and high pass filters as kernel of convolutional layer and will do the downsampling by using stride operator. But I am not sure whether that is right approach for facial emotion classification.
- Badges
- Science topic