Image Sentiment Analysis Integrating Global And Local Features

Nowadays,with the continuous improvement of network bandwidth,the function of social platform is more and more powerful.Compared with words,people are more inclined to use pictures or videos with more information to share their lives and emotions.Therefore,the emotional information contained in image data can also help people better grasp the public opinion tendency of an event or expand the selection range of keywords in image retrieval.Compared with the emotional analysis model which only focuses on the overall emotional characteristics of the image,how to use the emotional characteristics of different regions in the image to analyze the emotional tendency of the image has become the focus of research.But how to extract the local emotional features accurately? How to determine that the local emotional features of the extracted image play a positive role in the judgment of the emotional tendency of the image itself? These problems limit the development of image sentiment analysis model.To solve these problems,the main work of this paper includes the following points1.An emotion classification model that fuses global image emotion features with local emotion features is constructed,and the overall image emotion features are extracted by using a residual neural network,and the original feature extraction network in Faster-RCNN is improved by using an FPN network to extract emotion features in local regions of the image more effectively.2.Image emotions are classified into four categories according to the strength of polarity:positive,weakly positive,weakly negative,and negative,and an image emotion classification model based on the strength of emotion polarity is constructed based on the Res Net-101 residual neural network and the attention mechanism network.3.Through the weighted fusion of two image sentiment classification models,Image Emotion Recognition Model Based on Emotional Polarity and Holistic and Local Features is given,thus weakening the influence of sentiment uncertainty in some regions of the image and the classification error of each part of the model on the final sentiment classification,and making the model more accurate for image sentiment classification.4.Sentiment classification experiments were conducted on the FI image sentiment dataset and the MVSO image sentiment dataset,and the results showed that the proposed model achieved better classification results on these two datasets compared with other image sentiment classification models.