10 class problems with 324 features - am I recognition class based on multiple features, class based on the pattern of one feature? I am trying to balance 324 features for a 10 class problems, but it just seem humanly impossible, 
The problem i am studying consist of recognizing character , each observation consist of 324 features in the form pixels extracted from a 18x18 grid.  Should each pixel be seen as different  feature, or could one just turn the 324 feature into one feature, and thus easily compute the things -  Or would that increase error. 
So as the title states.
am I recognition class based on multiple features, class based on the pattern of one feature?
It should be said that the range of each pixels 0 -1. 
 A: With your data, each pixel can be treated as feature. So each sample consist of 324 numeric features (e.g. range [0,255]). You could now use those original features directly in your model, or you could do some more feature derivation first. The latter is likely good idea: you know your samples represent characters, therefore using e.g. Principal Component Analysis (PCA) is most probably going to help here. 
If I interpreted it correctly, PCA is one way to obtain what you asked for anyway:

[...] one just turn the 324 feature into one feature, and thus easily compute the things. 

PCA computes the "main components" of your data, which then can be used to express your samples as linear combination of these main components. This way PCA is often used for dimensionality reduction by discarding less important PC features. You can likely express your samples in much less than your original 324 features this way - though you will not be able to reduce your data to a single feature as still get good results.
