Abstract: In order to bridge the semantic gap between visual features and sentiments and reduce the impact of sentiment irrelevant regions in the image,this paper presents a novel visual sentiment prediction method by integrating visual semantic embedding and attention mechanism.Firstly,the method employs the auto-encoder to learn joint embedding of image features and semantic features,so as to alleviate the difference between the low-level visual features and the high-level semantic features.Secondly,a set of salient region features are extracted as input to the attention model,in which the correlations between salient regions and joint embedding features can be established to discover sentiment relevant regions.Finally,the sentiment classifier is built on top of these regions for visual sentiment prediction.The experimental results show that,the proposed method significantly improves the classification performance on testing samples and outperforms the state-of-the-art algorithms on visual sentiment analysis.

Key words: Visual sentiment prediction, Visual semantic embedding, Attention mechanism, Salient regions detection