I have a magnitude spectrum of (say) the letter "T". I want to extract features from this spectrum so I can classify it later. Can you suggest me any methods to do that?
feature design is often quite difficult. you could look into wavelets, PCA/LDA, fitting polynomials to your spectrum, spectral entropy, simply summing several bins that tend to be descriptive of your "T", filtering your 't' first then doing some of the aforementioned things.
really, this is one of the most difficult parts of designing a classifier. you'll need to decide what you want to discriminate between and design features that allow you to separate the classes.
Comments
really, this is one of the most difficult parts of designing a classifier. you'll need to decide what you want to discriminate between and design features that allow you to separate the classes.