Speakers
Mr
Piotr SzczepańskiMr
Karol ZielińskiMr
Albert Ziółkiewicz
Description
The research focused on classic image captioning based on a coder-decoder structure, where the coder encodes the image features. At the same time, the decoder produces a caption – a phrase describing the image content. We investigated the decoder part by testing multiple convolutional-neural-network-based backbones – feature extractors. This investigation aimed to find the optimal encoder, i.e., one that maximizes text generation metrics BLEU_1-Bleu_4, CIDEr, SPICE, and METEOR. Moreover, we worked on optimizing beam-search parameters used by the decoder to generate alternative phrases. Our research proves that an optimal choice of model’s hyperparameters increases caption generation efficiency.
Primary authors
Prof.
Marcin Iwanowski
Mr
Mateusz Bartosiewicz
Mr
Piotr Szczepański
Mr
Karol Zieliński
Mr
Albert Ziółkiewicz
