We present a high-level synthesis method called derivation from emulation results. A vision algorithm is first emulated in real time on a general image processing computer, then a chipset implementing the algorithm is automatically derived from the computer's configuration through multiple optimizations. The optimizations' correctness guarantees that the chipset will perform exactly as during emulation. A first version of the derivation software implementing low level optimizations has been developed; its output is a structural VHDL description. We present our first results and discuss our current research.