Big Five Personality Assessment Method Based On Visual Features And Adversarial Learning

Personality is the sum of the internal tendencies and psychological characteristics of an individual’s behavior in social adaptation to people,things,and oneself,etc.It describes a stable psychological state,rather than emotions that can change constantly within a short period of time.Personality assessment is widely used in vocational aptitude tests and entrance psychometric tests,and can help companies or employers determine a person’s stable emotional disposition in a short period of time when recruiting talent.The traditional Big 5 personality assessment relies on personality scales,and the results are obtained by the subjects’ own answers to a series of pre-designed personality test questions,which are in line with their own personality characteristics.This can cause serious interference with test results.With the development and advancement of deep learning technology,various Big Five personality assessments based on deep learning methods have emerged,while the existing personality assessment methods based on purely visual features are lacking in accuracy,and it is difficult to achieve high prediction accuracy for all five personality traits,and multimodal personality assessment methods are difficult to meet the demand for highquality input from multiple channels in practical applications.To address this situation,this paper proposes a personality assessment method based on purely visual features,optimizing the extraction process of visual features and combining it with adversarial learning to obtain a quantitative assessment of a person’s Big Five personality through videos of his or her daily life,as follows:First,a multiple coarse and fine-grained loss structure of the visual gaze direction estimation method is designed and implemented,and the trained gaze estimation model is used to obtain the gaze direction of the tester in the interval frames of the video,and further obtain the heat map of the tester’s gaze distribution and its gaze sequence characteristics,which reflect the overall distribution characteristics of the tester’s gaze in a certain time and the gaze time sequence characteristics,respectively.Secondly,the video picture frames are modeled by Time Sformer,and a gender discriminator is designed and trained jointly with the feature extractor,combining the idea of adversarial learning,so that the feature extractor tries its best to extract features that are not related to gender but to personality.Finally,all the extracted visual features are fused and their Big Five personality scores are predicted.The experimental results show that the average prediction accuracy of the method is 91.96% in the ECCV 2016 Personality Challenge dataset,which exceeds all previous visual feature-based Big Five personality assessment methods.In summary,for the inefficiency problems of the traditional personality scale based personality assessment methods,as well as the low accuracy of the existing deep learning based personality assessment methods and the difficulty in obtaining high-quality multimodal inputs that meet the requirements,the personality assessment method based on purely visual features and combined with adversarial learning proposed in this paper effectively solves these problems.