Research On Interactive Learning Based Semi-supervised Labeling Scheme And Action Recognition

Interactive learning is of great research significance in the field of computer science,and the main idea of which is to solve intricate problems by utilizing both human perception ability and computer accurate calculation ability.Interactive learning is widely used in fields such as robots and intelligent algorithms.However,it is not yet reached a clear conclusion regarding the design of interactive systems,in other words,how to design a manmachine cycle to complete the interactive process.Aiming at this problem,this paper proposes an interactive learning framework,which is applied in the fields of weakly supervised labeling and motion recognition.In the field of weakly supervised labeling,a dimension reduction failure case is detailed discussed first,which is cased by improper selection of dimension reduction subspace.Furthermore,the problem could be solved by interaction,after which the subspace could be decided again.Inspired by this idea,this paper proposes a visual interactive labeling system under the guidance of interactive learning framework.System design,algorithm design,and visual effect design are illustrated with the perception theory.The interactive system could be used in weakly supervised labeling field and weakly supervised model training applications.In terms of semi-automatic labeling,a labeling experiment is conducted on a Rubik's cube dataset,which achieves 99% labeling accuracy.As for supervised model training,handwritten numbers and facial emotion recognition experiment are conducted and both of which achieves remarkable performance with the current classification result under the system's semi-automatic labeling for model training.Among them,the performance of handwritten digital dataset achieves better classification accuracy rate than the adversarial generation network algorithm under labelrare cases.In the field of motion recognition,it caused a trouble that the current motion recognition algorithms is strongly relate with background features instead of the motion itself.This problem is particularly obvious in competitive sports field,under which the characteristic is shown as a single background,fast movement transitions,and complex movement types.Under the guidance of interactive learning framework,this paper design an interactive segmentation system,and by the help of it we propose a figure skating action dataset FSD-10.The collection process,labeling information,segmentation process,and action discriminant analysis of the dataset is analyzed in this paper.Next,this paper designs a key frame-based long-term segmentation neural network to evaluate the dataset.Considering the shortcomings of the existing network,the neural network is devoted to the core issue that how to extract key frames.Combining the understanding of motion characteristic,the network calculate the key frame by utilizing HPS parameter.The human dispersion is evaluated according to the HPS parameter,and key frames are extracted through the dispersion curve.In the experiment,compared with other networks,the network achieved excellent classification accuracy on the FSD-10 dataset.In addition,comparative experiments are also processed,which shows that FSD-10 is more sensitive to the action itself than the UCF-101 dataset.Based on the above analysis,the interactive learning framework and its application in the field of semi-automatic labeling and motion recognition are mainly illustrated in this paper.The main contributions of this paper are as follows.(1)This paper proposes an interactive learning framework combining related predecessor work and interactive learning related applications,which specifically illustrates the details of interaction in the human-machine loop.In the field of weakly supervised labeling,this paper designs a visual interactive learning system,which was finally applied to semi-automatic labeling and weakly supervised model training,and achieved excellent results compared with existing schemes.(2)In the field of action recognition,through an interactive segmentation system,this paper presents a competitive sports dataset FSD-10 with insensitive backgrounds,fastaction transitions,and complex types,which is of great significance for athletic action recognition.(3)In the motion recognition algorithm,this paper designs a key frame-based timing segmentation network.This network has achieved an excellent motion classification effect by designing a key frame extraction algorithm compared with the existing network structure.