How to select discrete cosine transform coefficients as a feature vector? I need to use DCT on frames of videos as a feature vector to train a Feed-forward artificial neural network, but the problem is the large number of coefficients (in thousands or so). How can I choose some of these coefficients and still capture the majority of information for my image (frame)?
 A: This depends on what you consider to be relevant information, which in turn depends on what kind of image features you are interested in.
The coefficients of a DCT describe progressively finer-grained structure in an image. So if you're interested in relatively large-scale features of your image, such as moving objects in the foreground, you are likely only interested in the low-frequency coefficients. On the other hand, if you are interested in small-scale features, such as the texture of materials, the high-frequency components are much more relevant. The figure below (from Wikipedia) shows the various components of the DCT, ranging from low-frequency in the top left, to high frequency in the bottom right.

As a quick and dirty approach, it's most likely that the information you're looking for is encapsulated in the first few low-frequency coefficients. You can try setting a cut-off $K$ and only accepting coefficients $X_{i,j}$ where $i^2 + j^2 < K$. The appropriate value of this cut-off can be determined either by trial and error, or by cross-validation.
