Image emotion classification technology based on polarity-aware attention mechanism

Abstract

Image emotion classification technology has been widely applied in fields such as social media. How to accurately extract emotional information from numerous images has become an essential issue in computer vision. To achieve high-precision image emotion classification, this study proposes a polarity-aware attention mechanism for both negative and positive coarse-grained sentiment classifications. Among them, polarity perception refers to the perception and classification of the binary polarity (positive/negative) of image emotions, while supporting fine-grained emotional intensity evaluation. At the same time, a classification network combining polarity-aware attention module and Transformer is introduced to achieve fine-grained emotion classification containing eight emotions. The results showed that the proposed polarity-aware attention mechanism classification method had an average accuracy of 95.88%, an average recall rate of 94.94%, and an average F1-value of 95.62% on the Twitter I binary sentiment classification dataset. In the ArtPhoto eight emotion classification datasets, the fusion of polarity-aware attention module and Transformer classification network achieved the highest classification accuracy of 98.23% and the lowest accuracy of 96.14% for the eight emotions. The highest classification accuracy of the weighted spatial context network was only 89.01%. The results indicate that the proposed image sentiment classification technique has achieved significant performance advantages in both coarse-grained and fine-grained sentiment classification tasks. The study provides important methodological support for applications in multiple fields including social media and sentiment computing.