Image Describing Based on Bidirectional LSTM and Improved Sequence Sampling

Motivated by great performance gained by Recurrent neural network applied on machine translation, people began to pay attention to image describing with related deep learning methods. Recurrent neural network can not remember long term information but Long-Short Term Memory(LSTM) can handle this well. However, the LSTM applied on image describing to predict sentences in previous literature [1] can only train and inference in the single direction. In fact, the words in a sentence not only relates to the context before but also later. In the paper, we propose a Bidirectional LSTM, it can generate sentences in both forward and backward direction with more richer information. Besides, we also improved sampling sentences. We conducted experiment on three datasets: Flickr8K, Flickr30K and MSCOCO datasets and our proposed models outperform related models.