Research On Multi-marker Learning Based On Marker Correlation

Single-label learning methods are often used to deal with the problem where each object associated with only one label.At present,the research on single-label learning is relatively mature,however,in the real world,objects often have multiple semantics,using multiple labels to describe an object at the same time may be more meet the actual needs,and existing single-label learning methods may not be able to effectively handle the above problem.Therefore,the multi-label learning framework is proposed to deal with the problem that an instance is related to multiple labels simultaneously.In order to improve the classification performance of the model,how to extract and utilize label correlation and learn through label-specific features are the current research hotspots of multi-label learning.In the past years,there have been many studies focusing on facial expression recognition.And it has a wide range of applications in the fields of intelligent robots,human-computer interaction,and synthetic animation.In real life,human faces usually express richer emotions.Multi-label learning algorithms can analyze multiple emotions expressed by human faces at the same time.Therefore,in this thesis,we have conducted in-depth study and research on the multi-label learning method based on the correlation of the labels and the learning of label-specific features,and achieved a multi-label facial expression system utilizing the proposed multi-label learning algorithms.Firstly,this thesis proposes a multi-label classification algorithm based on global and local label correlations.Most existing multi-label learning methods use either the global label correlation that is shared by all instances or the local label correlation when they exploit the label correlation.They assume that instances in different clusters should have different label correlations.In this chapter,we propose a novel multi-label learning method that utilizes both global and local label correlations to provide more comprehensive label information for the learning process.We have conducted extensive experiments on various fields of multi-label data sets to verify the effectiveness of the algorithm proposed in this chapter.The comparison results demonstrate the effectiveness of the proposed algorithm.Secondly,this thesis proposes a multi-label learning algorithm that joints label-specific features and correlation information.In this chapter,we propose a novel multi-label learning method,which also takes into account the learning of label-specific features and correlation information during the learning process.Firstly,we learn a sparse weight parameter vector for each label based on a linear regression model,and then we can extract label-specific features based on the corresponding weight parameters.Secondly,we constrain the label correlation directly on the output of the model,rather than on the corresponding parameter vectors that conflict with the learning of label-specific features.Finally,we exploit correlation among samples through sparse reconstruction relation.From the comparison results,it is significantly better than most existing multi-label learning algorithms.Thirdly,this thesis achieves a facial expression recognition system based on the proposed multi-label learning algorithms.At present,facial expression recognition has a wide range of application scenarios,while traditional facial expression recognition methods can only recognize one expression on a face image,and cannot effectively express people's complex and delicate emotions in daily life.The multi-label learning method can predict multiple expressions at the same time to more accurately reflect the true emotions of people.Therefore,we apply two multi-label learning algorithms proposed in this paper to achieve a facial expression analysis system.