Comparative Analysis between InceptionResnetV2 and InceptionV3 for Attention based Image Captioning

Generating a caption from an image is a challenging task that involves combining the concepts of computer vision and natural language processing which has motivated the researchers to model an automated process and in recent times multitude of approaches have been introduced which specifically target this problem. This paper solves the problem of learning the relationship between image and the language by establishing an entity-object relationship that maintains continuity in the caption. This is achieved by using an attention mechanism which allows the decoder to target specific regions at different time steps. To understand the image content, we use an encoder which extracts information about the different objects and the relationships between them. This paper evaluates and compares two models that incorporate two different encoders i.e., InceptionResnetV2 and InceptionV3, both having attention unit along with LSTM as the decoder. Experiments on the MS-COCO dataset show that the models generate sensible and accurate captions in a majority of cases.