Research On Image Discrimination Recognition Method Based On Body Features

With the development of social media,people can easily share images on the Internet.Some images always contain discriminatory information intentionally or unintentionally.The spread of discriminatory images is harmful to the construction of a harmonious and stable network environment.Identifying the discriminatory image effectively is a key method to block its spread,and has important research significance and application value.This paper carried out the research works on the model and algorithm of image discrimination recognition based on leveraging person body feature,including the construction of discriminatory image data sets,semantic modeling of discrimination relation and multi-features fusion method for recognition.The details are as follows:To solve the problem that there is no relevant discrimination data set at present,this paper constructs a well-annotated discrimination image data set,which contains 12287 images,obtained by twelve person characters in six scenes,eleven discrimination action categories and seventeen discrimination relation categories.The rationality of this data set is verified through experiments.Aiming at the feature of emotional interaction between people in discrimination images,a relation-based multi-task optimized learning model is proposed.This model uses the semantic transformation method to construct the relationship semantic space between people.In addition,multi-task loss module and feature enhancement unit are proposed to make the model output the action categories and the relationship categories at the same time.Our model uses the advanced semantic information of person actions to assist the prediction of discrimination relations.By comparing with the baseline algorithms,the effectiveness of this model for recognizing discrimination image is proved.In view of the importance of the local information of the person characters in the discrimination emotion judgment of image,this paper proposes a recognition method combining the local action information and global information of the person character,and taking into account the advantages of the relationship model and the traditional image recognition model in the network structure design.It guarantees to integrate the details of the action of the characters and the visual information of the emotional receiver.Through the comparative experiments,it proved that this method further improves the effectiveness of the algorithm model.