How to minimize the effect of pixel ovelapping in classifying images?

I am trying to classify some set of images into different classes using feature extraction.Color is one of the distinguishing features which i have considered.As some pixels of an image correlating with other image, the images are being identified wrongly

How can the Classification be improved by minimizing the effect of correlation among pixels of images?

Thanks

[Image Analyst]

Too general. Be more specific. It would take to long to explain this in a completely general context.