In this project, we will apply a Fisher LDA from a set of images of faces, each of size 256*256 in RGB format.
Dataset : jaffe
We reconstruct the original data by using K basis vectors obtained from LDA.
Below images show reconstructed images of one person for k=1, 6, 29):
Plot the MSE between the original and reconstructed images in terms of number of eigenvectors:
The problem of applying Fisher LDA on the dataset is that calculating the inverse of the covariance matrix of classes in high-dimensional data sets is very expensive. Also, if we have a large number of outliers in the dataset, A lot of noise is added to the average data, so lda cannot perform as expected.