Just read pixel, if its opaque add transformation to that coordinate(local texture coordinate) and then check sprite position + transformed pixel against geometry that you want.
Still I disagree the need of pixel perfect collision detection. Pixel perfect collision just means that you divide your object to size*size number of rects and bruteforce all possible collisions. Non naive implementation would use just as many rectangles that are needed for accurate collision detection. Bounding boxes are your friend learn to use them right.
Pixel perfect collision might work good enough for object vs static objects and only when you don't need collision forces but it's just not scalable or offers anything that would not be possile with other means.
I personally use Auriel Ribbon's box2d editor. http://code.google.com/p/box2d-editor/
I take my game object image and draw outlines and rest is magic. I would suggest you to think something similar in future.