Best RNN+CNN combination in image captioning

With the rapid development of deep learning, more and more scholars and researchers focus on image captioning, CNN+RNN is one of the greatest ways to solve image captioning. However, there is insufficient detailed research about which combination of CNNs and RNNs performs better for the CNN+RNN model. So the point of this article is to preliminary fill in the gaps in this section(Only the flicker8k dataset was considered). This article will mainly begin this study through a series of quantitative experiments and illustrate briefly the elaboration of the corresponding CNN pre-trained model and RNN structures. This essay will significantly compare, in terms of the model's loss function, the combination of mainstream pre-train CNN models (DenseNet201, ResNet50, VGG16, InceptionV3) and two types of RNN (LSTM, GRU) based on the filcker8K dataset. Finally, the conclusion of this essay is that the DenseNet structure is better than other structures for image capture and LSTM is worse than GRU in generating text.