Fine-Grained and Semantic-Guided Visual Attention for Image Captioning

Soft-attention is regarded as one of the representative methods for image captioning. Based on the end-to-end CNN-LSTM framework, it tries to link the relevant visual information on the image with the semantic representation in the text (i.e. captioning) for the first time. In recent years, there are several state-of-the-art methods published, which are motivated by this approach and include more elegant fine-tune operation. However, due to the constraints of CNN architecture, the given image is only segmented to fixed-resolution grid at a coarse level. The overall visual feature created for each grid cell indiscriminately fuses all inside objects and/or their portions. There is no semantic link among grid cells, although an object may be segmented into different grid cells. In addition, the large-area stuff (e.g. sky and beach) cannot be represented in the current methods. To tackle the problems above, this paper proposes a new model based on the FCN-LSTM framework which can segment the input image into a fine-grained grid. Moreover, the visual feature representing each grid cell is contributed only by the principal object or its portion in the corresponding cell. By adopting the pixel-wise labels (i.e. semantic segmentation), the visual representations of different grid cells are correlated to each other. In this way, a mechanism of fine-grained and semantic-guided visual attention is created, which can better link the relevant visual information with each semantic meaning inside the text through LSTM. Without using the elegant fine-tune, the comprehensive experiments show promising performance consistently across different evaluation metrics.