Perturbed Proximal Descent to Escape Saddle Points for Non-convex and Non-smooth Objective Functions