Boosting image caption generation with feature fusion module

Image caption generation has been considered as a key issue on vision-to-language tasks. Using the classification model, such as AlexNet, VGG and ResNet as the encoder to extract image features is very common in previous work. However, there is an explicit gap in image feature requirements between caption task and classification task, and has not been widely concerned. In this paper, we propose a novel custom structure, named feature fusion module (FFM), to make the features extracted by the encoder more suitable for caption task. We evaluate the proposed module with two typical models, NIC (Neural Image Caption) and SA (Soft Attention), on two popular benchmarks, MS COCO and Flickr30k. It is consistently observed that FFM is able to boost the performance, and outperforms state-of-the-art methods over five metrics.