Pixel-perfect 2D collision is much more complex than bounding-box collision, ESPECIALLY if the objects are able to rotate or scale. The general method is to have a bitmask for each object, which is then shifted to match their relative positions and then ANDed together - if two objects overlap, this will result in a bitmap that contains a "1" for every pixel that overlaps. If the objects are able to rotate and scale, that means the corresponding bitmasks have to be rotated and scaled as well - which often means writing a software rasterizer just to deal with this (although it's possible to make use of the GPU to assist with this).

(Another option is to store the left and right extents of the object for each scanline - but this method only works with vertically-monotone objects, and is not practical with rotating objects.)

The simplest option is give each object a collision shape defined by a set of several bounding boxes (this is the technique used by 2D fighting games, because it provides more accurate collisions for complex shapes and allows different parts of the object to have different collision properties) or, even simpler, bounding circles (since circles are rotationally invariant, collision detection is no harder if the objects can rotate, and additionally, circle collision is incredibly simple - if the distance between the circles' centers is less than the sum of their radii, they're colliding).

It's possible to make use of hardware acceleration to assist with per-pixel collision - but I doubt there are libraries for this, which means you'll be writing the shader code yourself.