A Deep Reinforced Training Method For Location-Based Image Captioning

Neural encoder-decoder frameworks have been used extensively in image captioning. Recent research has shown that reinforcement learning can be utilized to train these frameworks directly on non-differentiable evaluation metrics. However, the captions generated by this method usually have limited grammaticality and readability. In this paper, we propose a novel model with the location-based mechanism which introduces the location information of each region in the image, and a combined training method that combines the cross entropy loss and reinforcement learning. We evaluate our model on four public benchmarks: Flickr8k, Flickr30k, MSCOCO and Image Chinese Captioning (ICC). Experimental results show that our model can improve the readability of the generated captions and outperforms the state-of-the-art methods across different evaluation metrics.