Text Embedding Bank for Detailed Image Paragraph Captioning

Existing deep learning-based models for image captioning typically consist of an encoder to extract visual features and a language model decoder, architecture that has shown promising results in single high-level sentence generation. However, only the word-level guiding signal is available when optimized features. The inconsistency between parallel extraction sequential text supervision limits its success length generated long (more than 50 words). We propose new module, called Text Embedding Bank (TEB), address this problem paragraph captioning. This module uses vector learn fixed-length feature representations from variable-length paragraph. refer as TEB. TEB plays two roles benefit performance. First, it acts form global coherent regularize encoder. Second, distributed memory provide whole model, which alleviates long-term dependency problem. Adding existing state-of-the-art methods achieves result on Stanford Visual Genome dataset.