A Multi-level Style Feature Model for Visual Sentiment Analysis
Jiayu Wang, Dan Xu, Hao Zhang · 2024
Visual sentiment analysis aims to enhance user experience by predicting user sentiments and understanding their interests. However, existing methods typically rely on global learning to obtain the final sentiment representation of images in a one-time manner. They fail to fully utilize low-level visual features such as lines and colors, as well as the multi-level characteristics of images. To address these limitations, we propose an improved visual sentiment recognition method based on the convolutional neural network. Our approach extracts global deep features using the deep residual network and then integrate the statistical information of feature maps to extract style feature as local shallow features of the image. We perform feature fusion to obtain multi-level features. Specifically, by using a fusion function we fuse the style features from various levels into emotional feature representations. Additionally, we successfully apply style transfer techniques to the field of visual sentiment analysis in our study and validate their effectiveness through experiments.