I would go with something like sift features (or a non patented variant thereof if you plan commercial usage). You analyze a first picture and extract the cente

I would go with something like sift features (or a non patented variant thereof if you plan commercial usage). You analyze a first picture and extract the center part. You run a feature detector on each image and the can cheaply match (scale and rotation invariant) against each other - this gives rotation and scaling.