Learning figure-ground image segmentors by genetic programming

Figure-ground segmentation is an important image processing task that genetic programming (GP) has been successfully introduced to solve. However, existing GP methods use a homogeneous mixture of preprocessing and postprocessing operators for segmentation. This can result in inappropriate operators being connected, leading to poor performance and unnecessary operations in solutions. To address this issue, two new methods are designed to enable GP to conduct image preprocessing, binarisation and postprocessing separately. Specifically, the two methods introduce a strongly-typed representation (StronglyGP) and a two-stage evolution (TwostageGP) in GP respectively Results show that StronglyGP can evolve effective segmentors for the given complex segmentation tasks. However, TwostageGP currently performs poorly, which is likely caused by overfitting, which will be addressed in future work.