You're probably making your window size too small. What you're describing is one way that people do CBIR (Content Based Image Retrieval). Basically you get the mean and standard deviation of your image divided into tiles, then use them as feature vectors to find other images in a database that have a similar pattern. For example you could use it to find all pictures of a beach and sky similar to your sample image, but would not retrieve other images that had the same colors but laid out in a different pattern. Like it would get pictures with brown bottoms and blue tops but would not bring back a picture of a man in brown pants with a blue shirt. The example I saw divided the image up into 10 by 10 tiles and computed mean and stdev for hue, saturation, and value channels. So there were vectors of 6 hundred elements. You almost certainly don't need to have windows that are as tiny as 3 by 3. I'd say 100 tiles (10-by-10) is plenty. So then your feature vector is just a linear list of the means. Same for stdev. Just plug those vectors into the SVM machine. I don't have the Statistics Toolbox so I can't help anymore.
