Enhancing the Generalization Performance and Speed Up Training for DRL-based Mapless Navigation