Urban Acoustic Classification Based On Transfer Learning Feature Of Deep Convolution Neural Network

With the accelerating process of urbanization,activities such as construction,transportation,and social life will generate a lot of urban noise.Urban acoustic recognition plays a crucial role in noise monitoring,urban sound scene understanding,and noise source identification.Due to the low recognition rate of traditional hand-crafted acoustic features in urban acoustic recognition,this paper studies a high-precision urban environmental acoustic recognition method.The main work and achievements are as follows:1.An urban acoustic recognition method based on transfer features is proposed.First,the method uses the pre-trained deep Convolutional Neural Network(CNN)under the ImageNet dataset as the feature extractor to extract the deep features of the spectrogram.Then the transfer features are classified by a fully connected layer and the Softmax.This paper uses Inception-v3,ResNetl52,Inception-ResNet-v2 three kinds of networks for experimental comparison and verification.The feature fusion method is adopted to further improve the recognition accuracy of urban acoustic.2.A recognition method for transfer learning fusion features using Deep Belief Network(DBN)is proposed.First,the different DBN parameters are compared,including the number of hidden layer nodes,the number of Restricted Boltzmann Machines(RBM)training,and the number of DBN hidden layers,which affect the recognition results.Through the above experimental results,the DBN model with the highest recognition based on the deep CNN transfer learning fusion feature is obtained.Finally,the method is compared with the traditional acoustic recognition algorithm.The experiment proves the method that based on deep CNN transfer learning fusion using DBN classification can achieve 98.55%,which is superior to the traditional acoustic recognition method.