Improving Convolutional Image Captioning with Self-Critical Learning and Ranking Metric-based Reward