Quantifying the Impact of Complementary Visual and Textual Cues Under Image Captioning

Describing an image with natural sentence without human involvement requires knowledge of both image processing and Natural Language Processing (NLP). Most of the existing works are based on unimodal representations of the visual and textual contents using an Encoder-Decoder (EnDec) Deep Neural Network (DNN), where the input images are encoded using Convolutional Neural Network (CNN) and the caption is generated by a Recurrent Neural Network (RNN). This paper dives into a basic image captioning model to quantify the impact of multimodal representation of the visual and textual cues. The multimodal representation is carried out via an early fusion of encoded visual cues from different CNNs, along with combined textual features from different word embedding techniques. The resultant of the multimodal representation of the visual and textual cues are employed to train a Long Short-Term Memory (LSTM)-based baseline caption generator to quantify the impact of various levels of complementary feature mutations. The ablation study involves two different CNN feature extractors and two types of textual feature extractors, shows that exploitation of the complementary information outperforms the unimodal representations significantly with endurable timing overhead.