Research On Recognition Method Of Illegal Automobile Whistle Based On Deep Learning

As the number of automobile grows in the city,the pressure of transportation is increasing.The traffic pollution has become a significant problem.In these pollution,the car noise has had a great impact on people's life.To reduce the occurrence of phenomenon,corresponding laws have been issued by relevant law-enforcing department.But the benefits have been limited because of poor efficiency and drain on resource.Therefore,how to effectively locate and identify illegal whistle in city traffic has become a key issue to be studied in this field.The existing sound identification scheme about car whistle has some problems in the practical application process,such as low recognition efficiency,complicated calculation process,and excessive reliance on artificial review.In order to solve these problems,this paper applies the deep learning technology into car whistle recognition.As a generative model,depth belief network is used to identify car whistle.The feature of whistle signal is extracted in the convolution depth belief network and the feature is obtained better than the traditional feature classification.The existing methods and problems in the field of car whistle recognition are analyzed.A comparative study of existing recognition algorithms and the advantages of deep learning techniques over traditional recognition algorithms are made.This paper uses deep belief network as a generative model for car whistle recognition.The principle and extraction process of the traditional acoustic features on MFCC are described in detail.The obtained MFCC parameters are trained as the input layer of the DBN network,so that the deeper features in MFCC features are obtained.The soft-max classifier is added at the output layer of the network to complete the classification of the data.Through simulation experiments,it is demonstrated that different MFCC feature dimensions when inputed could make different recognition accuracy.And test sample lengths have an effect on the recognition accuracy.Compared with the recognition results of GMM model directly trained by MFCC feature,it proved the superiority of DBN in car siren sound.In order to obtain more representative acoustic characteristics than traditional features,this paper introduces convolution operation on the basis of deep belief network,and studies the feature extraction of whistle signals in using convolution depth belief network.The frequency characteristics of the car whistle signal are extracted.The frequency spectrum feature is used as the input of the CDBN network to train the network,and the CDBN characteristics are obtained.The GMM model is trained by the CDBN characteristics,and the data is identified.Through simulation experiments,it demonstrates that statistical values of feature parameters have an effect on the recognition results.And test sample lengths have an impact on the recognition results.At the same time,the identification accuracy of CDBN features is better than that of traditional MFCC,compared with the recognition results of GMM directly trained by model MFCC feature.