High-Quality Image Captioning With Fine-Grained and Semantic-Guided Visual Attention

The soft-attention mechanism is regarded as one of the representative methods for image captioning. Based on the end-to-end convolutional neural network (CNN)-long short term memory (LSTM) framework, the soft-attention mechanism attempts to link the semantic representation in text (i.e., captioning) with relevant visual information in the image for the first time. Motivated by this approach, several state-of-the-art attention methods are proposed. However, due to the constraints of CNN architecture, the given image is only segmented to the fixed-resolution grid at a coarse level. The visual feature extracted from each grid indiscriminately fuses all inside objects and/or their portions. There is no semantic link between grid cells. In addition, the large area stuff (e.g., the sky or a beach) cannot be represented using the current methods. To address these problems, this paper proposes a new model based on the fully convolutional network (FCN)-LSTM framework, which can generate an attention map at a fine-grained grid-wise resolution. Moreover, the visual feature of each grid cell is contributed only by the principal object. By adopting the grid-wise labels (i.e., semantic segmentation), the visual representations of different grid cells are correlated to each other. With the ability to attend to large area stuff, our method can further summarize an additional semantic context from semantic labels. This method can provide comprehensive context information to the language LSTM decoder. In this way, a mechanism of fine-grained and semantic-guided visual attention is created, which can accurately link the relevant visual information with each semantic meaning inside the text. Demonstrated by three experiments including both qualitative and quantitative analyses, our model can generate captions of high quality, specifically high levels of accuracy, completeness, and diversity. Moreover, our model significantly outperforms all other methods that use VGG-based CNN encoders without fine-tuning.