Spark-Based Semi-Supervised Ensemble Learning Video Semantic Concept Detection

With the rise of mobile Internet represented by smart portable devices,video semantic concept detection technology has become popular research direction today.Video semantic concept detection is also active in every corner of people's daily life such as Internet video,traffic safety,video surveillance and video medical.But because of the complexity,variety and noise of video content,video semantic concept detection research still faces enormous challenges.Based on reading a large number of domestic and foreign researches,a brief introduction about the background,significance and research status of video semantic concept detection is firstly made.Secondly,this thesis makes a brief introduction about semi-supervised learning,ensemble learning,and video semantic concept detection techniques.Thirdly,focused on the problems existing in domestic and foreign researches,this thesis proposes a semi-supervised ensemble learning video semantic concept detection algorithm based on pseudo-label confident selection and a Spark-based video semantic concept processing detection framework.Finally,in order to verify the availability of the proposed algorithm and framework,a Spark-based video semantic concept detection prototype system is proposed.The main research contents are as follows:(1)In order to solve the problems in video semantic concept detection that the insufficient labeled samples will seriously affect the performance of the detection and that the performances of the base classifiers in ensemble learning are improved deficiently due to noise in the pseudolabel samples,a semi-supervised ensemble learning algorithm based on pseudo-label confident selection is proposed.Firstly,three base classifiers on three different feature spaces are trained to get the label vector of the base classifier.Secondly,a method of weighted fusion of the error between the maximum and submaximal probability of a certain class to which sample belongs and the error between the maximum probability of a certain class to which the sample belongs and the average probability of the other classes to which the sample belongs is introduced to get the label confidence of the base classifier,pseudo-label and pseudo-label confidence of samples are confirmed through fusing label vector and label confidence.Thirdly,samples with the higher degree of pseudo-label confidence are added to the labeled sample set,and base classifiers are trained iteratively.Finally,the trained base classifiers are integrated to detect the video semantic concept collaboratively.The selected pseudo-label in the algorithm can reflect the overall variation among the classes to which the sample belongs and other classes,and the uniqueness of the class,which can reduce the risk of using the pseudo-label samples,the experimental results show that the proposed method can effectively improve the accuracy of video semantic concept detection.(2)In order to solve the problem of low efficiency of video semantic concept detection,a Spark-based video semantic concept processing detection framework is proposed.The framework includes the Spark-based video semantic key frames extraction framework,the Spark-based video semantic features extraction framework,the Spark-based video semantic model training of semisupervised ensemble classifier framework and the Spark-based video semantic classification and detection framework,and it can process concurrently key frames extraction,features extraction,video semantic model training and video semantic classification and detection in memory.In this thesis,comparative experiments are carried out in different experimental environments.The experimental results show that the framework proposed can significantly improve the speed of key frames extraction,features extraction,video semantic model training and video semantic classification and detection and improve operational efficiency of video semantic concept detection.At the same time,this thesis also summarizes the impact of the number of available CPU cores and the available memory size of the cluster on the operational efficiency of the framework,and factors that constrain the efficiency of the framework.(3)In order to verify the usability of algorithm and framework proposed,The Spark-based video semantic concept detection prototype system was designed and implemented by using technologies or environments,such as Ubuntu16.04,HDFS,Python,Sklearn and OpenCV.The prototype system has many characteristics,such as simple interface,easy to use and user-friendly,so the usability of the proposed algorithm and framework is fully validated.