[Purpose/significance] Currently, researches on sentiment analysis of online users in online public opinion
events focus on texts, texts combined with images or videos, lacking the discussion on images. Additionally, there is a
lack on the visual sentiment analysis fusing multiple visual semantics features. [Method/process] We draw the idea
from multimodal fusionand employ the thought of multimodal fusion as the theory guidance of multiple visual semantics
feature fusion. We design four different visual sentiment analysis models based on feature-level fusion, intermediate fusion, decision-level fusion and hybrid fusion. The designed models are basedon VGG 19and Xception pretrained on
the ImageNet. [Result/conclusion] We conduct empirical experiments on image datasets in online public opinion
events and compare the proposed models with baseline models. Experiments show that the model based on decision-level fusion is superior. To hence the interpretability of the model, we visualize the outputs from the convolutional lay ?
ers in the visual sentiment analysis model.