Acoustic Scene Classification And Sound Event Detection Based On Deep Learning

Acoustic scene classification(ASC)and sound event detection(SED)are key technologies in many fields,such as multimedia analysis and retrieval,audio monitoring,intelligent driving.In this thesis,the ASC and SED methods based on deep learning are investigated.The main work and contributions of this thesis are as follows:(1)A method for ASC based on audio feature augmentation(AFA)is proposed.This thesis mainly discusses the impacts of audio feature augmentation on the performance of ASC.Specifically,the similarities and differences of the features of two channels are obtained by performing mean and difference operations on the features of two channels audio data.Harmonic percussive source separation(HPSS)is applied to the audio spectrogram to obtain augmented audio features.When evaluated on the databases of DCASE2016 and DCASE2017,the accuracies obtained by the method based on audio feature augmentation are85.8% and 69.9% respectively,and are higher than that obtained by the methods without data augmentation.In addition,the proposed method is compared with other acoustic scene classification methods,and is better than other methods.(2)A method for detecting abnormal sound event of road based on deep feature fusion is proposed.First,the deep autoencoder network(DAN)is used to transform three shallow features,i.e.,mel frequency cepstrum coefficient(MFCC),Bark filter bank(BFB)and Gabor filter bank(GFB),to deep features.Next,the three deep features above are concatenated and then transformed to a new deep feature by a DAN.Finally,the new deep feature is fed into a long short term memory network(LSTM)for decision of ASC.The experimental results show that the proposed feature obtains detection accuracy of 92.15% and F score of 91.32%without adding background noises to the audio signal,and outperforms other features for abnormal sound event detection.After adding background noises to the audio signal for obtaining different signal-to-noise ratios(SNRs),such as 20 dB,10 dB,0 dB and-10 dB,the proposed deep feature obviously outperforms other features,and has a strong anti-noise ability.In summary,this thesis proposes a ASC method based on audio feature agumentationand a method of road abnormal sound event based on deep feature fusion.The performance of the proposed methods is analyzed by experiments from various aspects.The reliability and effectiveness of the proposed method are verified under various experimental conditions.