Research On The Group Feature Selection Algorithm And Its Application In The Early Recognition Of Flame Video

In recent years,fire recognition technology has developed from sensor identification to the video image recognition.Starting from the static features and dynamic features for fire smoke,fire early recognition method which is based on video image has got certain improvement in accuracy,however the accuracy of the fire video identification still exists a lot of room to improve when facing to the video image contains a lot of disruptors.As the fire video early recognition is essentially a data classification problem,so only starting from the features of fire smoke are not able to effectively improve the prediction precision of the algorithm.In this paper,according to the features of fire in the video identify in extraction,on the basis of a large number of flame features,the key is to improve the back-end classification algorithm,thus improving the precision of fire recognition.The main work and contributions are as follows:First of all,in order to find the relationship between features,pick out the optimal feature subset,this paper proposes a new group feature selection algorithm based on the optimization of feature subset discrimination.The algorithm will maximize intra-group feature correlation and inter-group feature discrimination as large as possible at the same time,and then an optimal group structure is determined adaptively.Experimental results on UCI data sets demonstrate that the proposed algorithm can identify the inherent group structure in features and has better classification performance compared with the state-of-the-art algorithms.Secondly,as the high dimensional fire smoke feature selection is easy to fall into dimension disaster problem,In this paper,we propose a Group Feature Selection Based on L_1/2 Regularization method to discover the feature groups and get less features.The feature coefficients are constrained by a L_1/2 regularization operator,which urges the coefficients of feature to be zero.Besides,we constructs a distance criterion of features,which will make the same set of coefficients in the same group as possible,and proposes a joint optimization algorithm based on ADMM,at last achieve the purpose of finding the group structures.Experimental results on synthetic and Hill_Valley datasets demonstrate that the proposed methods can find the inherent group structure in features.Finally,in order to solve the present flame video features less data and most data need to be set threshold in early recognition threshold shortcomings,we developed a software of fire video early identification.using the real tunnel fire video,it can automatically extract ten typical flame smoke characteristics.In a addition,Using the proposed algorithm to construct an early fire recognition model,experimental results show that these features can meet the requirements of the proposed algorithm,and validate the proposed algorithm in the early recognition of flame video having higher classification accuracy.