Weakly supervised discriminate enhancement network for visual sentiment analysis

Several methods employ weakly supervised technology to highlight the visual sentiment information in images, so as to improve the performance of sentiment analysis. However, the over-focusing of location technology leads to the neglect of the information of some discriminative sentiment regions. In this work, we propose a W eakly S upervised D iscriminate E nhancement N etwork (WSDEN) for visual sentiment analysis to construct a prediction framework of image sentiment. To be specific, firstly, the proposed WSDEN learns sentiment maps with a weakly supervised technology, thus it significantly reduces the annotation burden and encourages features to convey more sentiment information. Secondly, to solve the problem of overemphasis on local information mentioned above, Discriminate Enhancement Map is constructed by spatial weighting and channel weighting, which combines the sentiment map to perform sentiment enhancement on the final deep features of images. Extensive experiments on three benchmark databases show that the proposed method outperforms the state-of-the-art methods of visual sentiment analysis.