Research On Abnormal Sound Feature Extraction And Classification In Public Places Based On Spectrogram

The security of public places for social stability and harmony, people’s lives andproperty is of great significance. Currently, the video surveillance system has becomethe principal means of access to information of all kinds of abnormal event. Abnormalsound recognition based on audio surveillance in public places as a complement andsupplement of the video surveillance, it can effectively reveal the occurrence ofabnormal events. Therefore, Abnormal sound feature extraction and classification inpublic places is of significant practical value and academic significance. Featureextraction of abnormal sound mostly follow the traditional methods of speech signalprocessing such as Mel-Frequency Cepstrum Coefficient （MFCC）, Linear PredictionCepstrum Coefficient （LPCC） and so on. The abnormal sound in public places includesthe speech signal, such as screaming and non-speech signals such as explosions, gunfireand crackle. In addition, there is the interference from sound of car horns, conversationvoices, footsteps and low-frequency atmospheric noise interference in public places. So,the traditional feature extraction method based on speech processing is obviouslyinsufficient. Based on the above analysis, and through the analysis of abnormal soundspectrogram, the thesis proposed that firstly public places abnormal sound signals areconverted to the spectrogram, secondly2D-Gabor filter characterization is used for thecharacteristics of the sound spectrum; and then Stochastic Non-negative IndependentComponent Analysis （SNICA） is used for extracting abnormal sound spectrogramcharacteristics, and finally using the Sparse Representation Classification （SRC） methodfor classification. The main content of the work is described as follows:①Pretreatment of the abnormal sound in public places.1) The characteristics ofthe background noise in public places behave that energy spread evenly, short-termenergy is relatively stable; however, abnormal sound signal behaves impulse, short-termenergy focusing on, and prominent in the background noise. So this thesis proposes adual-threshold method of short-term energy for extracting abnormal sound clips fromthe background noise in public places.2) because the background noise of public placesin line with the distribution of Symmetry Alpha-Stable（S S）, In this thesis, theminimum average p norm （LMP） is proposed to remove background noise of the publicplaces. By comparison with the wavelet threshold denoising method, the effectivenessof the proposed method is verified. ②Through the analysis of the principle of the formation, the pitch frequency andspectrogram of abnormal sound in public places, this thesis asserts that significantlydifference of direction and subtle of abnormal sound spectrogram time-frequencystructure, can reveal the nature of the abnormal sound signals, but also have gooddiscriminative. However, there is no obvious time-frequency structure in thespectrogram of the background noise in public places which distributed evenly. So,effective feature can be extracted from spectrogram for the classification, at the sametime suppressing background noise. Because of the above analysis, this thesis proposedthat firstly abnormal sound signals are converted to the spectrogram, and then, in orderto better tap the amount of information of the spectrogram time-frequency structuralcharacteristic, time-frequency structure is described based on2D-Gabor filter. Theseworks facilitate follow-up feature extraction and classification of abnormal soundspectrogram.③The characteristics of the abnormal sound spectrogram is similar within classand significantly different between of classes. this paper establishes the global sparserepresentation model based on public places abnormal sound spectrogram, and theabnormal sound feature extraction and classification method is proposed by the model.SNICA is used for extracting sparse feature-based based on2D-Gabor time-frequencystructural description. This feature can be a better reflection of high-order andnon-negative characteristics of the sound spectrogram, while effectively overcome theoverlap effect in the spectrogram. And then SRC based onl₁Norm minimization is usedfor the representation of the identified signal in the dictionary, which is used to classify.By comparison with experiments in traditional methods, the effectiveness of theproposed method is verified.