Research On The Interestingness Understanding And Prediction Method Of Images And Videos

With the rapid development of computer^ vision and artificial intelligence,the increasing popularity of digital multimedia,the continuous improvement of people's pursuit of life has brought about a massive increase in data,but the quality of existing data is mixed.Through the research and prediction of interestingness,it can help people to complete the retrieval of relevant information efficiently,and has a positive effect on advertising promotion,video summary and on-demand broadcasting.Based on the machine learning method,this paper constructs the corresponding prediction model from the perspective of the two-classification prediction of images and videos interestingness,so that the computer can simulate the human perception mode and automatically complete the two-classification task of images and videos interestingness.In order to describe the concept of interestingness and convert it into a computable problem,this paper builds an image interestingness prediction framework.Firstly,the framework identifies three important cues that describe the interestingness,namely unusualness,aesthetics,and general preferences.Each cue consists of different types of features,of which unusualness consists of outliers and familiarity features,aesthetics consists of arousal,texture,color,complexity and shape features,general preferences consists of local features and scene descriptors.Secondly,feature fusion for the same type features is performed by discriminant correlation analysis or multi-set discriminant correlation analysis.Finally,the simple multiple kernel learning method is used to classify the interestingness of images.The experimental results show that the interestingness prediction framework proposed in this paper can capture the interestingness information of the image comprehensively,achieve higher classification accuracy and have good prediction performance.For the two-classification task of video interestingness,in order to solve the problem that the static feature ignores the video dynamic information,and the video visual information cannot be fully described.In this paper,the combination of static features and dynamic features is used to characterize the interestingness,and the video interestingness is classified by the AdaBoost classifier.The static feature extracts color histogram,SIFT,HOG,Gist,and LBP features for the video frames.Based on the video frames,the three orthogonal planes of XY,XT and YT are modeled,and the local binary pattern features on the three planes are extracted respectively,and they are connected in series as dynamic features to describe the spatio-temporal information of the video.Experimental results show the dynamic and static features combination method used in this paper makes up for the lack of video dynamic information,and has a good performance on video interestingness classification.