Analyzing The Performance Of Automatically Captioned Images
Pondsingh Jackie, G Shrine, Sravan Yadav Eadala · 2014
Images are automatically captioned in two stages, Content Selection and Surface Realization. Content Selection is based on probabilistic model that suggests the keywords for the image and description in the Article. Surface realization techniques which determines how to verbalize the chosen keywords. In Content Selection the keyword assumption is based on probabilistic image annotation model that the images and the surrounding text are generated by a latent variables or topics. The surface realization emphasizes the caption generation in extractive and abstractive methods. Whereas in extractive method a sentence is generated from the keywords obtained through the former method discussed above, the abstractive method creates the caption which may be from the word based or phrase based caption generation approach. So the idea here is to use phrases for surface realization. The approach is to analyze the performance of automatically captioned images using probabilistic approach. Indeed the output of the abstractive model is obtained using phrases rather than words.