Strong Three Way Decision Theory And The Application In Computer Vision

Three Way Decision(TWD)is the expansion of two way decision.In practical applications,people often can not immediately make rejection or acceptance due to incomplete and uncertain information.Therefore,it is necessary to use TWD for data mining in this situation.The positive domain,negative domain and boundary domain of TWD are interpreted as accepting action,rejecting action and not making action,and then the decision rules can be obtained.According to the minimum risk principle,TWD gives a reasonable loss function condition,that is,the cost of accepting a good decision is lower than the cost of delaying decision,and delaying decision cost is lower than the cost of rejecting decision.How-ever,the relationship between the cost functions of TWD is weak,which can not solve the data imbalanced problem and cost-sensitive problem.Therefore,how to choose the appropriate cost function according to the practical problems is a worthy question to study.TWD can be used in the computer vision to deal with the uncertainty information problems and cost-sensitive issues.For instance,imbalanced data problem would be inevitably faced in image classification.In the video anomaly detection,the cost of classifying abnormal behaviors as normal behaviors is much higher than the cost of classifying normal behaviors as abnormal behaviors.How to apply TWD to computer vision is a problem which we concerned about.The main work is as follows:?Aiming to solve the problem of the weak relationship between the cost functions of TWD,this paper put forward a theory called Strong Three Way Decision(STWD)and its dual form(D-STWD).At the same time,this paper define a decision factor? and analyze the relationship between the boundary domain and ?,besides,the re-lationship between STWD,D-STWD and decision thresholds ? and ?.The example of decision table explains the difference between STWD and General TWD(GTWD).When dealing with imbalanced problems,STWD can pay more attention to the smaller scale class;when dealing with cost-sensitive issues,high cost class will be emphasized.?Based on the proposed STWD,this paper proposes a multi-classification method called STWD-Ensemble and applies it to the image classification problem.STWD mainly solves the data imbalance problem in the multi-class classification.More-over,this method makes full use of the three domains in the GTWD:the positive domain,the negative domain and the boundary domain,which retains possibilities of one image belongs to different classes as far as possible.The proposed method makes the final decision by integrating multiple weak classifications.The experi-mental results show that our proposed multi-classifier algorithm based on STWD achieves higher precision,recall and F-Score compared to GTWD and some other machine learning models like Bayesian Network,Random Forest and so on.?This paper applies the proposed STWD to video anomaly detection,which mainly solves the cost-sensitive problem.Firstly,extract the low-level and high-dimensional features of frames by the optical flow method.Next,reduce the dimension and extract the semantic features through the topic model,and then use the proposed STWD to train a classifier to classify the actions.Experiments show that STWD is much better than the GTWD and other machine learning models like Naive Bayes and Random Forest,especially in the recall and F-Score of abnormal behaviors.