The Research On Image Sentiment Classification Method Based On Multiple Visual Object Learning

Image sentiment analysis is a complex problem.The main difficulty lies in the diverse causes and variegated expressions of emotions.When an object appears alone,it may not show a discriminatory attitude,but it conveys discriminatory emotion when it appears simultaneously with other related objects in the image.Discrimination as a social injustice phenomenon is harmful to the social development and unity,which affects the social relationships.Currently,techniques such as convolutional neural network and deep learning have good performance in image classification and object detection,but there is still much space for improving the accuracy and fineness of emotion classification.The relationships between image emotions and objects in the image and the impact of multiple objects on image emotion expression have not been intensive studied.This paper proposes an analysis method based on Discrimination-Sensitive Object Classes and Attributes(DSOCAs)for a human-specific and high-level semantic emotion-discrimination.The main work is as follows:(1)For image discrimination emotion detection and determination problems,analyze the factors that constitute discrimination and build dataset.Aiming at the relevance of multi-visual objects that constitute image discrimination,a multi-visual objects and multi-attribute discriminatory emotional learning framework is proposed.(2)The fusion of object detection and migration learning technology obtain discriminatory sensitive objects.In view of the multiple attributes of the same discriminatory sensitive object,the multi-task deep learning neural network is constructed by utilizing the relevance of the different attributes of the same object to acquire the multiple attribute features and improve the learning efficiency.(3)Aiming at the correlation between image discrimination and multiple visual objects,a discriminatory classification method based on conditional random fields was proposed to improve the validity and accuracy of image discrimination judgment.(4)Based on the existing Emotions in Context Database(EMOTIC)and self-constructed discriminatory sentiment dataset,the discriminatory emotions are subdivided into 5 categories(ageism,sexism,racism,language discrimination,body language discrimination).Experiment is conducted to integrate discriminatory sensitive target objects,their object attributes and locations.The experimental results show that the precision,recall and F1 of our method are obviously improved compared with the baseline method.