I am not sure i understand the concept of CNN and why it is so good at image processing. I always be told that CNN is good at finding unique features, thereby good at classifying stuff.. But does that make sense?
I guess i understand how the conv. layer extract features and pooling layer maximized an window in from the convolution layer, but how are the these features used for classification in the fully connected layer. As far I know is the fully connected layer similar to a neural network, so what basically it does is classification using the features extracted from the prior steps, in which a training would make sense here, but then again, one wouldn't extract unique features but extract already known features, and train a neural network based on that?... If that is the case?... What is the benefit of using CNN? If it is basically a NN with a preprocessing/feature extraction step?
How come is CNN so good image processing, if it require one to know what features are being extracted, and train a neural based on that... It basically boils downs to a regression problem?
To state the question in a way so that it makes sense with the title - How come is CNN a thing, when the thing it does combine feature extraction (known) + Neural network (training)?
