An improved content-adaptive mesh-generation method for image representation