Hi to all,

I'm currently working on a face recognition project. The database i use is ORL database. I'm currently using SOM for feature extraction, then use K-NN for classification.

I got a relatively good recognition rate, about 96%. Recognition rate is the total number of images recognised correctly / total number of test images. However, my SOM's quantization error seems to quite bad, about 3.5

So i tried normalizing the dataset before using it to train the SOM.
The ave.quantization error then drops to about 0.3 but the recognition rate falls to 40%! Before this, i used PCA also.. and applied normalization and the results were worse then if i did not use normalization. I'm pretty sure my normaliztion method is correct because i got from a reliable online source.

Can anyone help me by explaining why the recognition rate drops when i use normalization? Is it to do with the database? Why is it that SOM's ave.quantization error is big when the recognition result is good?