Comparison of Image Encoder Architectures for Image Captioning

Image captioning is a fascinating and challenging task combining two interesting fields in computer science - Natural Language Processing and Computer Vision. Several approaches have been tried for this task. Many of these approaches are based on a combination of Convolutional Neural Networks(CNN’s) as encoders and Recurrent Neural Networks(RNN’s) as decoders. This paper compares two CNN-based encoders – (Vector Geometry Group) VGG16 and InceptionV3, which are used as encoders for encoding the input image features. The proposed research work uses the popular Flickr8k dataset and also compares among two different loss functions used with each of the above architectures- Categorical Cross Entropy loss function and Kullback-Leibler Divergence. Our main goal is to study the effect of the image encoder and the loss function on the image captioning task while keeping all other parameters the same.