• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Anti?noise sound recognition based on energy?frequency feature

    2015-12-03 05:18:22ZHOUXiaominLIYing
    智能系統(tǒng)學(xué)報(bào) 2015年5期

    ZHOU Xiaomin,LI Ying

    (College of Mathematics and Computer Science,F(xiàn)uzhou University,F(xiàn)uzhou 350108,China)

    Anti?noise sound recognition based on energy?frequency feature

    ZHOU Xiaomin,LI Ying

    (College of Mathematics and Computer Science,F(xiàn)uzhou University,F(xiàn)uzhou 350108,China)

    In the natural environment,non?stationary background noise affects the animal sound recognition directly.Given this problem,a new technology of animal sound recognition based on energy?frequency(E?F)feature is proposed in this paper.The animal sound is turned into spectrogram to show the energy,time and frequency characteristics.The sub?band frequency division and sub?band energy division are carried out on the spectrogram for extracting the statistical characteristic of energy and frequency,so as to achieve sub?band power distribution(SPD)and sub?band division.Radon transform(RT)and discrete wavelet transform(DWT)are employed to obtain the important projection coefficients,and the energy values of sub?band frequencies are calculated to extract the sub?band frequency feature.The E?F feature is formed by com?bining the SPD feature and sub?band energy value feature.The classification is achieved by support vector machine(SVM)classifier.The experimental results show that the method can achieve better recognition effect even when the SNR is below 10 dB.

    animal sound recognition;sub?band power distribution(SPD);sub?band frequency feature;radon transform(RT);energy?frequency(E?F)feature

    http://www.cnki.net/kcms/detail/23.1538.tp.20151014.1432.002.html

    Ecological environmental sound events contain a wealth of information closely related to the human sur?vival environments.The relevant information of envi?ronmental sound can be obtained by sound processing technology.Thus the environment sound recognition technology can be used to analyze and survey environ?ment,predict the environmental change,detect spe?cies intrusion,monitor the animal life habit and protect them.The research on environmental sound is of great significance to protect and develop the environment.However,in real situation,there is a large amount of non?stationary noise who seriously disturbs the main sound signals.Thus,how to identify the main sound signals from the noisy environment is the emphasis and difficulty in sound signal processing.

    Currently,the signal processing technologies are mostly concentrated in the speaker recognition and mu?sic classification,while environmental sound process?ing is relatively less.Since the environmental sound is drawing more and more attention,some scholars haveput forward some new recognition technologies.For ex?ample,Chou et al.[1]segmented the bird song to obtain the syllables and used Wavelet transformation on Mel?frequency cepstral coefficients(WMFCCs),then clas?sified audio with BP neural network.Chen et al.[2]de?veloped an automatic frog call recognition system based on the combination of a pre?classification method of syllable lengths with a multi?stage average spectrum(MSAS)method to recognize 18 classes of frog calls.Chu et al.[3]used the matching pursuit(MP)algo?rithm to obtain effective time?frequency features,and combined MP features with MFCC features to obtain higher accuracy.Khunarsal et al.[4]used spectrogram pattern matching with the classifiers of neural network and k?nearest neighbor(k?NN)to recognize the short time environmental sound.You et al.[5]employed time encoded signal processing and recognition(TESPAR)to identify the environmental sound events,which used the TESPAR coding method to produce simple data structures,and then used the archetypes technique for classification.Dennis et al.[6]presented the sub?band power distribution(SPD)image with k?NN classifier to recognize the sound events.

    However,there are still some disadvantages in theabove methods.The MFCC feature is sensitive to the change of signal,so the background noise may serious?ly degrades the recognition accuracy.Syllable length pretreatment is susceptible to the influence of overlap?ping sound signals.MP algorithm calls for a large a?mount of calculation while searching the optimal atoms in over?complete dictionary.It is different to use spec?trogram pattern matching to design the relevant factors,as it has no noise resistance.The recognition effect of TESPAR is not satisfying even though the method is of less complexity.The SPD image presented by Dennis et al.[6]can effectively describe the characteristic of the spectral power distribution,but the method must be applied in a special and known noise environment.In view of the above questions,combining with the advan?tage of SPD image,this paper proposes a feature ex?traction method based on energy?frequency(E?F).E?F refers to calculating the sub?band energy density distri?bution and the sub?band energy value of the sound sig?nal based on the spectrogram.Then the sound signals are described using the energy distribution characteris?tic and the sub?band energy value feature.

    1 E?F feature

    The details of the recognition framework are shown in Fig.1.It can be seen that feature extraction is the core part of the classification system.The effectiveness of feature parameters may directly affect the classifica?tion result.Therefore,a novel method of E?F feature,based on the time?frequency characteristics of sound signal,is proposed in this paper.The method applies SPD image to depict the distribution of sub?band ener?gy and describe the relationship between sub?band fre?quency and sub?band energy level.In the meantime,the sub?band energy value feature is used to represent the relationship between sub?band frequency and ener?gy value.Combining with the two features mentioned a?bove,the effective E?F feature of different sound sig?nals can be extracted.

    Fig.1 Animal sounds recognition framework

    1.1 SPD feature

    The SPD image is designed to represent the distri?bution of spectral power in each frequency sub?band o?ver time.So it needs to divide the frequency and the energy of spectrogram into several sub?bands respec?tively.Then the corresponding energy level of each pix?el is calculated in spectrogram.The result can be ob?tained to describe the spectral power distribution[6].

    1.1.1 SPD image

    The SPD describes the energy distribution in each frequency sub?band over time.Assuming that sound signal x(n)is translated into spectrogram P(t,f).For the convenience of processing,it can be turned P(t,f)into P(f,t).The frequency can be divided into sub?band based on the frequency range.Assuming that the spectrogram can be divided into Fbsub?bands,the fre?quency range ofeach sub?band frequencyAfter sub?band division,the spectrogram can be repre?sented as P(fb,t) ,

    To unify the normalization results in the range of[0,1],the negative adjustment can be carried out as fol?lows:

    Then use the sub?band energy division to calculate the density distribution of sub?band energy to get its SPD image.The SPD is based on the density distribution of G(fb,t),so the distribution can be chosen by a non?parametric approach that based on the histogram.Ac? After sub?band division,the spectrogram can be repre?sented as P(fb,t),cording to the energy distribution,the number of sub?band can be set as Zb,then the SPD can be written as follows:

    where tmaxis the number of time samples in the seg?ment,1zb(G(fb,t))is the indicator function,which e?quals one if and only if the G(fb,t)belong to zbth bin and is zero otherwise,namely,

    The raw probability distribution HR(fb,zb)can be got according to the above processing,in which 0≤HR(fb,zb)≤1.And it can be found that most of the information is contained within a small region of the dy?namic range based on the SPD.Besides,the distribu?tion density values of the sub?band energy vary with the sound signals.To enhance the contrast of the raw SPD,it can be used Eq.(6)to obtain the enhanced SPD image to obtain important information.

    1.1.2 Radon transform(RT)and discrete wavelet transform(DWT)

    RT can not only describe the directional feature and local feature of the image to reduce the feature pa?rameters of the image but also have advantage of noise resistance.Assuming that F′(x,y)is corrupted by the noise F(x,y),then the distance between the two sig?nals before projection can be defined as

    After the RT,the distance between the two signals has become as W2,

    After turning the spectrogram into SPD image,RT coefficient can be extracted.RT in the SPD image serves as one method to calculate the projective trans? form at certain directions,which is to calculate the line integrals of image along the parallel lines.The RT of the SPD image HR(fb,zb)is defined as Refs.[7?10],

    where Rθ(r)is new image resulting from the RT;r=fbcos θ+zbsin θ;θ∈[0,180)is the angle between the distance vector and the fbaxis;?[HR(fb,zb)]de?notes RT of the 2?D function HR(fb,zb);and δ(x)=is the Dirac function.RT of the SPD im?age is shown in Fig.2.In the SPD image,the value and color of each point on the Z axis denotes the corre?sponding energy distribution probability,and the paral?lel lines represent the projection direction.The energy of linear integral in each parallel line can be calculated and the corresponding value can be reflected on the Z axis to obtain the projection results.

    In this paper,the result of RT at a certain angle is a one?dimensional serial Rθi(r′),where r′=1,2,…,N,N is the number of projection points,θi∈{θ1,θ2,…,θM},M is the number of projection angles.Hence the corresponding projection matrix can be ex?pressed as

    Using the above method,the Radon projection co? efficients at different angles can be got.However,if the projection coefficients are used directly to identify spectrogram,there will be a large amount of calcula?tions as well as a relatively low recognition rate.Aiming at the problem in hand,the feature parameters need be selected by some conditions.Since the DWT has the characteristics of frequency separation,noise elimina?tion and improvement of signal resolution,it can be a?dopted to extract the important parameters from the Ra?don projection coefficients.In this paper,the db3 wavelet function is applied to realize the decomposition and to obtain the transform coefficients.Wavelet coeffi?cients are divided into low frequency part and high fre?quency part,in which case the low frequency part de?notes the main structure information of signals and the high frequency part denotes signal details.It can beknown that the low frequency can reveal more about the features of signals,so the low frequency part of DWT is selected as the feature parameters.

    Fig.2 RT of the SPD image

    1.2 Sub?band feature

    The SPD image represents the distribution of sub?band energy,and the pixel in SPD image corresponds to the probability of the sub?band energy level.Howev?er,the SPD feature cannot show the specific value of sub?band energy.Since the energy value of sub?band frequency varies with the animal sound,the sub?band energy value can be calculated to distinguish the ani?mal sounds.Therefore,the paper applies the sub?band division to extract specific energy value,which is used to represent the sub?band feature.

    The principle of sub?band division is to set the range of frequency,and base on the range to divide the spectrogram into multiple sub?bands.The spectrogram is divided into N bands with equifrequency based on Eq.(1),so the spectrogram P(t,f)is divided into P(t,fk).The number of frequency in each sub?band isfrequency band.In sub?band, the sliding window can be set based on the characteris?tics of frequency band,and the energy characteristics of the spectrogram within windows can be extracted.The features of the sub?band are adopted to describe the frequency characteristics of spectrogram.Due to the stability of frequency volatility of animal sounds,the effective description of sub?band feature can improve the accuracy of classification.

    The paper extracts the SPD image of sound signal and applies it to represent the probability distribution relationship between sub?band frequency and sub?band energy.Then,calculating the energy value of sub?band by combining the sub?band frequency feature,so as to obtain the distribution feature and value feature of the E?F characteristics.

    2 Sound recognition system framework

    Sound recognition is a standard pattern recognition process.The system framework includes three stages:pre?processing,feature extraction and recognition.In pre?processing,an improved multi?band spectral sub?traction(IMSS)method is used to reduce the noise of sound signals and remove the silence sections after IMSS.The key of feature extraction is to denote the sound signal after pretreatment by citing the E?F fea?ture proposed above in this paper.In recognition stage,SVM classifier is used to train and recognize the above features and then obtain the results.The following part will show the concrete steps of pre?processing and fea?ture extraction in detail for specific animals.

    2.1 IMSS to de?noise

    The noise estimation of traditional multi?band spectral subtraction is calculated during the period of silence.It can obtain the approximate noise estimation in a stable and high SNR environment,but in the non?stationary and low SNR environment,it can not esti?mate the noise accurately.

    In general,as real noise is non?stationary,the traditional multi?band spectral subtraction fails to accu?rately estimate the actual noise,which will lead to in?complete noise reduction,and even distort the original sound signal.So estimating the noise accurately is the core problem of multi?band spectral subtraction.To solve this problem,the paper proposes the IMSS based on the controlled recursive averaging noise estimation method[11].IMSS can not only accurately estimate the noise power spectrum of the sounds,but also divide the sounds into multiple frequency bands and conduct spectral subtraction targeting to the affection differences of each band by noises so as to receive a better de?noi?sing effect.

    Assuming that the additive noise signal d(n)is uncorrelated with clean sound signal s(n),the corrup?ted sound y(n)can be expressed as

    where y(n)is divided into overlapping frames by the application of a window function and analyzed by using the short?time Fourier transform,

    where k is the frame index,f the frequency bin index,N1the length of frame,M frame shift,and h(n)a win?dow function.So Eq.(11)can be written as follows:

    Assuming that H0(k,f)and H1(k,f)indicate sound absence and presence respectively in the kth frameofthefthfrequencybin,λd(k,f)=denotes power spectrum of noise sig? nal,and λ^d(k,f)is power spectrum estimation of noise signal.In general,a temporal recursive smoothing method will be adopted to perform the noise estimation during the period of sound absence,that is,

    where αd(0<αd<1)is a smoothing parameter which is set as 0.85.

    Supposing that p(k,f)defined as P(H1(k,f)|Y(k,f))denotes the conditional signal presence proba?bility,for the case of H0(k,f)and H1(k,f)in Eq.(14),the noise estimation of the next frame can be obtained by full probability events as follows:

    If α~d(k,f)denotes αd+(1-αd)p(k,f),Eq.(15)can be expressed as

    Then,the noise power spectrum estimation can be ob?tained through α~d(k,f)and recursive averaging of sound power spectrum.Basing on the multi?band spec?tral subtraction proposed in Ref.[12],the recursive averaging is applied to calculate the noise power esti?mation and replace the original noise power estimation,thereby improving the traditional multi?band spectral subtraction in order to cope with noise reduction better.

    There are kinds of algorithms used in the compare experiment,namely,IMSS,the popular multi?band spectral subtraction[12],wavelet de?noising method(WDM)[13]and speech enhancement algorithm with leading?in delay(SEAWLD)[14].The compare experi?ment uses 57 class animals with 570 sound signals to reduce noise in different noise environments under dif?ferent SNRs.From Table 1,it can be seen that the IMSS has a better de?noising effect,and reduce the non?stable noises effectively.

    Table 1 The result of noise reduction in different environ?ments with different de?noising method dB

    2.2 Remove silence

    The feature extraction of sound signal would di?rectly affected by the existence of silence parts in the de?noise signal,so the silence must be removed before extracting feature.The paper uses the short?time energy threshold to remove the silence.No segmentation of o?verlapping frames is conducted to obtain each frame s^i(j),where i is frame index,and j the sample point within a frame.Then the short?time energy can be ob?tained as follows:

    where L is the length of a frame.The short?time energy of all frames can be calculated by Eq.(17),then the total energy and average energy can be obtained.The threshold is set based on the average energy multiplied by parameter a(a is set based on the actual situation,which is set as 0.05 in this paper).If the short time energy is smaller than threshold,the signal is consid?ered as silent and removed,and the effective sound signal s′(n)is finally obtained.

    Framing,windowing and short?time Fourier trans?form for sound signal s′(n)will be carried out to obtain S′(k,f)by Eq.(12),and the corresponding energy spectrum density function(power spectrum function)P(k,f)can be described as

    P(k,f)is a two?dimensional nonnegative realfunction.Abscissa will be expressed with time t,repla?cing the original frame k,and with f as the ordinate,the spectrogram is constructed by gray?scale which is expressed by the values of the corresponding P(t,f)[15].

    2.3 E?F feature extraction

    After the pre?processing,sub?band energy divi?sion is carried out to achieve the sub?band energy prob?ability distribution.Then RT is applied to extract the directional feature and the local feature of the SPD im?age,and combining with DWT to obtain the important parameters.After the energy distribution feature is ob?tained,the sub?band frequency division is carried out to extract the sub?band energy value feature.And then,the sub?band energy distribution feature and sub?band energy value feature are combined to form the E?F fea?ture set of sound signal.

    According to the generation principle of SPD im?age,the sub?band division and sub?band energy divi?sion are carried out on the normalized spectrogram to obtain the distribution of energy in each frequency's sub?band.The sparrow's spectrogram is shown in Fig.3(a),and Fig.3(b)represents the SPD figure corresponding to Fig.3(a).From Fig.3,it can be con?cluded that the frequency range of spectrogram is from 0 to 22 050 Hz,and the frequency is divided into 129 bins.The energy is divided into 100 bins.Pixel colors of the spectrogram can be expressed by energy value.Warmer tone of the pixel color indicates higher energy value,for example,the deep blue value is 0.The spectrogram shows that energy mainly concentrates in the mid?frequency domain,which is displayed in the SPD image.The pixel color of the SPD diagram means the frequency of corresponding sub?band energy,the higher the probability,the warmer the color tone.The animal sound spectrogram is different and the corre?sponding energy distributions are not the same,so the SPD image can effectively distinguish the sound type.

    The data of the SPD figure is large,and if the SPD figure is served as the feature vector,it may lead to large computational complexity and long training time.In order to effectively extract the main feature of the image,reduce the calculation and shorten the cal?culation time,the paper use RT to extract the direc?tional feature and the local feature of the SPD image.

    Fig.3 The sparrow's spectrogram and corresponding SPD image

    The RT result of each particular angle is a one?di?mensional serial as shown in Fig.4(a),which are the curves of the sparrow’s spectrogram after the Radon projection at 30°and 75°.It can be seen that the pro?jection coefficients are different at different angles and Radon projection curve can effectively describe the tex?ture and shape features of image.

    Taking 15°as a spacing,in the range of[0, 180),12 projection coefficientsat specific di? rections can be calculated,where θi∈{θ1,θ2,…,θ12},the corresponding values are

    [0 15 30 45 60 75 90 105 120 135 150 165]The DWT is applied after getting the projection coeffi?cients.In the experiment,the db3 wavelet function is adopted in DWT to transform the Radon projection co?efficients of each angle respectively.DWT divides the signal into high frequency and low frequency parts as shown in Fig.4(b).Since the low frequency can re?veal more about the feature of signal,the low frequen?cy part can be selected as the feature parameters.

    In order to effectively describe sub?band energy value feature of sound signal,the paper divides thespectrogram into multiple bins.Then the sliding win?dow can be applied to divide each band to calculate the energy value of each window as features.As shown in Fig.5,the left part of figure divides the spectrogram into 9 bins,and the right part denotes the third and fourth frequency band.In the experiment,the paper divides the spectrogram into 100 bins,sets the sliding window as the size of sub?band,and then calculates the energy values of sub?band as features.

    Fig.4 RT and DWT of sparrow sound's SPD image

    Fig.5 Sub?band division

    3 Experiments

    3.1 Experimental design

    Table 2 shows the 60 subclasses of 3 classes of ec?ological sounds,which include sounds of birds,mam?mals and natural noise from the Freesound data?base[16],freeSFX database[17],and are recorded by SONY ICD?UX512F recorder.Each subclass of sounds includes 30 segments with 2 s length,44 100 Hz sam?pling frequency and 16 bit digitization.

    In this paper,for each subclass of animal sounds,20 clean of sound segments are randomly selected as train samples.And the rest of 10 sound segments,which are mixed with different types noises with the SNRs of 0,5,10,20,30,40 dB,are experimented as test samples so as to acquire the animal sound recog?nition rates under different SNRs.

    In order to verify the effectiveness of the proposed method,the existing method in Ref.[18]and the popular MFCCs are selected to make a comparison.In Ref.[18],after carrying out the pre?processing of multi?band spectral subtraction,the anti?noise feature APNCC is extracted to feed into SVM for classification.

    It can be drawn from Table 1 that IMSS can effec?tively reduce the non?stationary noise and keep the mu?tation and detail part of the sound signal.Under the same situation,the IMSS can get better results.There?fore,two groups of contrast test were carried out in this paper.The first group is to extract E?F features,RT+DWT features,APNCC features,and MFCCs features respectively after de?noising by IMSS.Then SVM clas?sifier is used to classify them,and then the different recognition rates under noisy environments are com?pared in order to reflect the availability of E?F features.The second group is to experiment with the method of Ref.[18]and the popular MFCCs combined with SVM to reflect the practicability of the proposed approach.In this paper,the LIBSVM[19]package designed by Dr.Lin Chih?Jen of Taiwan University is used directly,and the RBF kernel is chosen experimentally.

    3.2 Results analysis

    Table 3 shows the recognition rates of four differ?ent methods in noiseless environment by the above ex?perimental design.It is observed that four kinds of methods can achieve high recognition rates in noiseless environment.

    Table 2 60 subclasses of 3 classes of ecological sounds

    Table 3 The recognition rate of four different methods in noiseless environment%

    However,in actual situation,there is no noiseless environment.To solve the problem,the paper applies the sound of water,waterfall and rain to simulate the real environment.And experiment is made with the noises which have different SNRs,to obtain the results.

    Fig.6 shows the results of the first group of exper?iment.It shows the recognition effects of four kinds of methods under three different noisy environments after de?noising by IMSS.It can be known that the recogni?tion effect of the method proposed in the paper is better than the other three methods when the SNRs are within the range of 0 to 40 dB.When the SNR is above 10 dB,the recognition rate of the E?F feature extrac?tion is higher than 90%.And it can be got that the RT+DWT can improve the recognition rate in low SNR en?vironment,but the recognition effect is not as good as the E?F feature in high SNR environment.Fig.6 shows that the recognition effect of the APNCC is not ideal and the MFCC can recognize the sound signals with high SNR,but is easily affected by the noise signal.

    The main reason for the above results is that E?F feature and RT+DWT feature are based on the image which is relatively insensitive to noise.Due to that E?F feature is based on the SPD image and the sub?band division spectrogram,it can describe the sound signal more effectively.While APNCC and MFCC feature ex?traction are based on the analysis of the signal frame,they are more sensitive to signal absence and signal ad?dition.Thus the change of signal may directly affect the recognition accuracy.

    Fig.6 Recognition accuracy under three kinds of noises

    Table 4 shows the results of the second group ex?periment.It can be seen that the average recognition rate of the method proposed in the paper is about 20% higher than that in the Ref.[18]and about 22%high?er than the popular MFCCs.Through the comparison of specific data,it can be got that the recognition rates of Ref.[18]and popular MFCC have increased when in high SNR environment,while in low SNR environ?ment,the E?F feature has obvious advantages and the recognition effect is better than the other two methods.

    Table 4 The recognition accuracy of three kinds of methods under different noisy environments%

    4 Conclusions

    For animal sound recognition in a noisy environ?ment,this paper adopts IMSS to reduce noise,and puts forward the E?F feature extraction method.The ex?perimental results show that the recognition accuracy of this method is obviously higher than popular MFCC and APNCC under non?stationary noise environment.The results of experiments show that the proposed method can effectively improve the recognition rate of animal sounds under low SNR environments,and has high noise resistance and robustness.The fact that the IMSS has affected the sound signal leads to the concentration of recognition rates on about 94%even in a high SNR environment.The next step of this research work is mainly to reduce the influence of the de?noising method for sound signals.

    [1]CHOU C H,LIU P H.Bird species recognition by wavelet transformation of a section of birdsong[C]//Symposia and Workshops on Ubiquitous,Autonomic and Trusted Compu?ting.Brisbane,Australia,2009:189?193.

    [2]CHEN W P,CHEN S S,LIN C C,et al.Automatic recog?nition of frog calls using a multi?stage average spectrum[J].Computers&Mathematics with Applications,2012,64(5):1270?1281.

    [3]CHU S,NARAYANAN S,KUO C C J.Environmental sound recognition with time?frequency audio features[J].IEEE Transactions on Audio,Speech,and Language Pro? cessing,2009,17(6):1142?1158.

    [4]KHUNARSAL P,LURSINSAP C,RAICHAROEN T.Very short time environmental sound classification based on spec?trogram pattern matching[J].Information Sciences,2013,243:57?74.

    [5]YOU Guanyu,LI Ying.Environmental sounds recognition u?sing TESPAR[C]//2012 5th International Congress on Im?age and Signal Processing(CISP).Chongqing,China,2012:1796?1800.

    [6]DENNIS J,TRAN H D,CHNG E S.Image feature repre?sentation of the subband power distribution for robust sound event classification[J].IEEE Transactions on Audio,Speech,and Language Processing,2013,21(2):367?377.[7]AJMERA P K,JADHAV D V,HOLAMBE R S.Text?inde?pendent speaker identification using Radon and discrete co?sine transforms based features from speech spectrogram[J].Pattern Recognition,2011,44(10/11):2749?2759.

    [8]JAFARI?KHOUZANI K,SOLTANIAN?ZADEH H.Rotation?invariant multiresolution texture analysis using Radon and wavelet transforms[J].IEEE Transactions on Image Pro?cessing,2005,14(6):783?795.

    [9]JADHAV D V,HOLAMBE R S.Radon and discrete cosine transforms based feature extraction and dimensionality re?duction approach for face recognition[J].Signal Processing,2008,88(10):2604?2609.

    [10]JADHAV D V,HOLAMBE R S.Feature extraction using Radon and wavelet transforms with application to face rec?ognition[J].Neurocomputing,2009,72(7/8/9):1951?1959.

    [11]COHEN I,BERDUGO B.Noise estimation by minima con?trolled recursive averaging for robust speech enhancement[J].IEEE Signal Processing Letters,2002,9(1):12?15.[12]KAMATH S,LOIZOU P.A multi?band spectral subtraction method for enhancing speech corrupted by colored noise[C]//IEEE International Conference on Acoustics Speech and Signal Processing(ICASSP).Orlando,USA,2002:4164?4164.

    [13]CHIEMENTIN X,KILUNDU B,RASOLOFONDRAIBE L,et al.Performance of wavelet denoising in vibration a?nalysis:highlighting[J].Journal of Vibration and Con?trol,18(6),2012:850?858.

    [14]LIU Xiang,GAO Yong.Speech enhancement algorithm with leading?in delay[J].Modern Electronics Technique,2011,34(5):85?88.

    [15]LI F Q,WAN H,HUANG J J.The display and analysis of sonogram based on MATLAB[J].Microcomputer Informa?tion,2005,21(10?3),172?174.(in Chinese)

    [16]Universitat Pompeu Fabra(UPF)[EB/OL].[2014?05?09].http://www.freesound.org.http://www.freesound.org.

    [17]FreeSFX.co.uk Team[EB/OL].[2014?05?09].http://www.freesfx.co.uk/soundeffects.

    [18]YAN Xin,LI Ying.Anti?noise power normalized cepstral coefficients in bird sounds recognition[J].Acta Electroni?ca Sinica,2013,41(2):295?300.(in Chinese)

    [19]CHANG C C,LIN C J.LIBSVM:a Library for Support Vector Machines[EB/OL].[2014?05?15].http://www.csie.ntu.edu.tw/~cjlin/libsvm.

    About authors:

    LI Ying was born in 1964.He re?ceived the B.S.degree in industrial auto?mation in 1989,the M.S.degree in sys?tem engineering in 1993 and the Ph.D.degree in computer science and technology in 2004,all from Xi’an Jiaotong Univer?sity,Xi’an,China.Since 2011,he has been a professor at the College of Mathematics and Computer Science of Fuzhou University,F(xiàn)uzhou,China.His current re?search interests are environmental sound recognition and infor?mation security.He has published more than 20 papers and 6 patents in these areas.

    TP391.4 Document code:A Article ID:1673?4785(2015)05?0810?10

    10.11992/tis.201408028

    :ZHOU Xiaomin,LI Ying.Anti?noise sound recognition based on energy?frequency feature[J].CAAI Transactions on Intelligent Systems,2015,10(5):810?819.

    Received Date:2014?08?21. Online Publication:2015?10?14.

    Foundation Item:Supported by the National Natural Science Foundation of China(No.61075022).

    LI Ying.Email:fj_liying@fzu.edu.cn.

    网址你懂的国产日韩在线| 九色成人免费人妻av| 97在线视频观看| av视频在线观看入口| 成人性生交大片免费视频hd| 国产精品一区二区免费欧美| 黄色一级大片看看| 一进一出抽搐gif免费好疼| 日韩中字成人| 一夜夜www| 伦理电影大哥的女人| 国产爱豆传媒在线观看| 99国产极品粉嫩在线观看| 深夜a级毛片| 午夜a级毛片| 亚洲综合色惰| 国产一区二区激情短视频| 免费看美女性在线毛片视频| 国产精华一区二区三区| 天天躁日日操中文字幕| 国产黄色视频一区二区在线观看 | 国产高清不卡午夜福利| 3wmmmm亚洲av在线观看| 综合色丁香网| 国产69精品久久久久777片| 小蜜桃在线观看免费完整版高清| 日韩欧美 国产精品| 国产一区亚洲一区在线观看| 国产成人影院久久av| 色尼玛亚洲综合影院| 黄色一级大片看看| 精品久久久久久久久亚洲| 欧美一区二区亚洲| 国产精品一二三区在线看| 中文字幕av在线有码专区| 一进一出抽搐gif免费好疼| 99在线人妻在线中文字幕| 春色校园在线视频观看| 国内久久婷婷六月综合欲色啪| 日韩一区二区视频免费看| 人妻制服诱惑在线中文字幕| 国产精品一区二区性色av| 亚洲av美国av| 99久国产av精品国产电影| 中文字幕人妻熟人妻熟丝袜美| 亚洲av二区三区四区| avwww免费| 欧美zozozo另类| 国产精品久久久久久久久免| 亚洲欧美日韩东京热| 亚洲精品在线观看二区| 香蕉av资源在线| 老师上课跳d突然被开到最大视频| 欧美极品一区二区三区四区| 联通29元200g的流量卡| av专区在线播放| 中文字幕av在线有码专区| 亚洲中文字幕日韩| 午夜a级毛片| 国产精品一区二区三区四区久久| 亚洲国产精品久久男人天堂| 国产一区二区在线av高清观看| 看非洲黑人一级黄片| 一边摸一边抽搐一进一小说| 乱人视频在线观看| 久久久久久国产a免费观看| av国产免费在线观看| 国产午夜福利久久久久久| 波多野结衣巨乳人妻| 欧美中文日本在线观看视频| 免费观看的影片在线观看| 亚洲七黄色美女视频| 国产亚洲91精品色在线| 久久中文看片网| 久久精品夜夜夜夜夜久久蜜豆| 22中文网久久字幕| 在线a可以看的网站| 欧美最新免费一区二区三区| 国产亚洲精品av在线| av黄色大香蕉| 99热这里只有精品一区| av女优亚洲男人天堂| 一个人免费在线观看电影| 久久久久久久久久黄片| 美女大奶头视频| 日韩欧美 国产精品| 你懂的网址亚洲精品在线观看 | 亚洲国产精品sss在线观看| 国产成人精品久久久久久| 黄色日韩在线| 婷婷亚洲欧美| 亚洲成av人片在线播放无| 国产精品久久久久久久久免| 好男人在线观看高清免费视频| 最近手机中文字幕大全| 亚洲国产精品成人综合色| 亚洲七黄色美女视频| 天美传媒精品一区二区| 亚洲国产欧洲综合997久久,| 99热只有精品国产| 丰满人妻一区二区三区视频av| 日韩欧美一区二区三区在线观看| avwww免费| 日本五十路高清| 久久精品国产自在天天线| 精品国产三级普通话版| 日韩中字成人| 中文字幕av成人在线电影| 人人妻,人人澡人人爽秒播| 国产真实伦视频高清在线观看| 精品人妻偷拍中文字幕| 丝袜喷水一区| 国产综合懂色| 亚洲天堂国产精品一区在线| 尤物成人国产欧美一区二区三区| 久久久久精品国产欧美久久久| 国产色爽女视频免费观看| 国产黄色视频一区二区在线观看 | 欧美绝顶高潮抽搐喷水| 国产精品亚洲一级av第二区| 国产高清视频在线播放一区| 一a级毛片在线观看| www.色视频.com| 嫩草影视91久久| 国产精品女同一区二区软件| 精品一区二区三区视频在线观看免费| 国产爱豆传媒在线观看| 日本免费一区二区三区高清不卡| 在线观看av片永久免费下载| 村上凉子中文字幕在线| 一区福利在线观看| 国产精品,欧美在线| 久久中文看片网| 哪里可以看免费的av片| 好男人在线观看高清免费视频| 深爱激情五月婷婷| 一夜夜www| 亚洲精品影视一区二区三区av| av在线播放精品| 日本免费a在线| 国产精品女同一区二区软件| 美女内射精品一级片tv| 日韩欧美精品免费久久| 亚洲欧美日韩高清专用| 国产美女午夜福利| 精品人妻一区二区三区麻豆 | 大型黄色视频在线免费观看| 成人三级黄色视频| 国产精品国产三级国产av玫瑰| 亚洲性夜色夜夜综合| 久久久久国内视频| 美女 人体艺术 gogo| 成年av动漫网址| 亚洲人成网站在线观看播放| 国语自产精品视频在线第100页| 一进一出抽搐gif免费好疼| 黄片wwwwww| 久久精品影院6| а√天堂www在线а√下载| 国产精华一区二区三区| 在线看三级毛片| 亚洲精品乱码久久久v下载方式| 午夜激情福利司机影院| 熟女人妻精品中文字幕| 一区二区三区高清视频在线| 18禁黄网站禁片免费观看直播| 日本与韩国留学比较| 精品一区二区免费观看| or卡值多少钱| 日日干狠狠操夜夜爽| 亚洲精品国产av成人精品 | 少妇熟女aⅴ在线视频| 久久精品国产亚洲av涩爱 | 天堂网av新在线| 搡老熟女国产l中国老女人| 免费观看在线日韩| 欧美日本亚洲视频在线播放| 一级毛片电影观看 | 欧美高清成人免费视频www| 日韩,欧美,国产一区二区三区 | 成人鲁丝片一二三区免费| 欧美最黄视频在线播放免费| 一级黄色大片毛片| 少妇人妻一区二区三区视频| 亚洲熟妇熟女久久| 国模一区二区三区四区视频| 欧美xxxx性猛交bbbb| 直男gayav资源| 欧美日韩国产亚洲二区| 熟女电影av网| 欧美日韩精品成人综合77777| АⅤ资源中文在线天堂| а√天堂www在线а√下载| 精品欧美国产一区二区三| 一卡2卡三卡四卡精品乱码亚洲| 午夜免费男女啪啪视频观看 | 欧美日韩乱码在线| 久久午夜亚洲精品久久| 亚洲一区高清亚洲精品| 精品人妻一区二区三区麻豆 | 亚洲成人中文字幕在线播放| 亚洲自偷自拍三级| 两个人视频免费观看高清| 三级毛片av免费| 又爽又黄a免费视频| 两性午夜刺激爽爽歪歪视频在线观看| 久久久久久九九精品二区国产| 我的女老师完整版在线观看| 中国国产av一级| 精品欧美国产一区二区三| 国产精品永久免费网站| 日韩成人伦理影院| 99久久久亚洲精品蜜臀av| 国产精品av视频在线免费观看| 日韩成人av中文字幕在线观看 | 午夜福利在线观看免费完整高清在 | 亚洲欧美清纯卡通| 亚洲精品乱码久久久v下载方式| 亚洲性久久影院| 欧美日韩一区二区视频在线观看视频在线 | 国产真实伦视频高清在线观看| 麻豆国产97在线/欧美| 亚洲欧美清纯卡通| 校园春色视频在线观看| 国产午夜精品论理片| 日韩一本色道免费dvd| 精品久久国产蜜桃| 好男人在线观看高清免费视频| 国产精品一区二区免费欧美| 小说图片视频综合网站| 亚洲自拍偷在线| 床上黄色一级片| 夜夜夜夜夜久久久久| 51国产日韩欧美| 国产精品电影一区二区三区| 国产一级毛片七仙女欲春2| 久久久成人免费电影| 偷拍熟女少妇极品色| 人妻少妇偷人精品九色| 99热这里只有是精品50| av福利片在线观看| 少妇丰满av| а√天堂www在线а√下载| 99久国产av精品国产电影| 日本黄色片子视频| 国产伦在线观看视频一区| 97碰自拍视频| 黄色视频,在线免费观看| 国产亚洲欧美98| 日韩成人伦理影院| av天堂中文字幕网| 精品熟女少妇av免费看| 国产精品国产三级国产av玫瑰| 免费观看的影片在线观看| 尤物成人国产欧美一区二区三区| 99久久九九国产精品国产免费| 少妇被粗大猛烈的视频| 超碰av人人做人人爽久久| 国产精品美女特级片免费视频播放器| 天堂网av新在线| 一进一出抽搐gif免费好疼| 男女边吃奶边做爰视频| 国产视频内射| 久久鲁丝午夜福利片| 免费av毛片视频| 日本欧美国产在线视频| 免费观看精品视频网站| 人妻夜夜爽99麻豆av| 国产精品日韩av在线免费观看| 国产成人aa在线观看| 国产伦精品一区二区三区四那| 91午夜精品亚洲一区二区三区| 欧美性猛交黑人性爽| 久久人人爽人人片av| 99久久精品一区二区三区| 身体一侧抽搐| 亚洲自偷自拍三级| 精品一区二区三区视频在线| 色5月婷婷丁香| 精品人妻视频免费看| 日本a在线网址| 色av中文字幕| 人人妻人人看人人澡| 女同久久另类99精品国产91| 亚洲精品成人久久久久久| 亚洲av.av天堂| 嫩草影院入口| 亚洲欧美清纯卡通| 亚洲一区高清亚洲精品| 亚洲精品乱码久久久v下载方式| 人人妻人人看人人澡| 国产精品久久久久久亚洲av鲁大| 激情 狠狠 欧美| 久久99热6这里只有精品| 中文字幕久久专区| 国产亚洲精品久久久久久毛片| 色在线成人网| 淫秽高清视频在线观看| 国产美女午夜福利| 免费观看精品视频网站| 此物有八面人人有两片| 99在线人妻在线中文字幕| 欧美色欧美亚洲另类二区| 熟妇人妻久久中文字幕3abv| 亚洲乱码一区二区免费版| 久久久久久大精品| 麻豆成人午夜福利视频| 99热这里只有是精品50| 国产亚洲精品久久久久久毛片| 日韩制服骚丝袜av| 国产人妻一区二区三区在| 国产伦精品一区二区三区四那| 寂寞人妻少妇视频99o| 国产国拍精品亚洲av在线观看| 国产美女午夜福利| 国产探花极品一区二区| 精品久久久久久久久亚洲| 亚洲av第一区精品v没综合| 女的被弄到高潮叫床怎么办| 国产黄片美女视频| 国产精品免费一区二区三区在线| 禁无遮挡网站| 国产精品野战在线观看| 日韩一本色道免费dvd| 日韩成人av中文字幕在线观看 | 天天躁夜夜躁狠狠久久av| 久久韩国三级中文字幕| 日韩精品中文字幕看吧| 69av精品久久久久久| 久久精品国产亚洲av涩爱 | 国产日本99.免费观看| 国产免费一级a男人的天堂| 日日摸夜夜添夜夜爱| 国产免费男女视频| 一个人免费在线观看电影| 在线播放无遮挡| 亚洲欧美成人综合另类久久久 | 69人妻影院| 午夜精品国产一区二区电影 | 欧美成人一区二区免费高清观看| 亚洲国产精品国产精品| 三级国产精品欧美在线观看| 欧美性猛交╳xxx乱大交人| 欧美高清性xxxxhd video| 久久精品91蜜桃| 一边摸一边抽搐一进一小说| av天堂中文字幕网| 99riav亚洲国产免费| 日本成人三级电影网站| 久久久久九九精品影院| 午夜福利在线观看免费完整高清在 | 毛片女人毛片| 精品一区二区三区人妻视频| 国产淫片久久久久久久久| 亚洲一区二区三区色噜噜| 亚洲最大成人手机在线| 日本黄大片高清| 在现免费观看毛片| 亚洲成人久久性| 久久久国产成人免费| 色哟哟哟哟哟哟| 好男人在线观看高清免费视频| 永久网站在线| 亚洲av免费高清在线观看| 亚洲一区二区三区色噜噜| 国产亚洲精品久久久com| 亚洲国产日韩欧美精品在线观看| 噜噜噜噜噜久久久久久91| 国产精品电影一区二区三区| 丰满乱子伦码专区| 亚洲,欧美,日韩| 视频中文字幕在线观看| 亚洲国产精品999| 国产亚洲91精品色在线| 下体分泌物呈黄色| 免费黄色在线免费观看| h日本视频在线播放| 亚州av有码| 国产日韩欧美视频二区| 国产男女超爽视频在线观看| 精品一区在线观看国产| 晚上一个人看的免费电影| 日韩av在线免费看完整版不卡| 人妻人人澡人人爽人人| 久久久欧美国产精品| 亚州av有码| 国产精品蜜桃在线观看| 免费观看av网站的网址| 成人18禁高潮啪啪吃奶动态图 | 嫩草影院入口| h视频一区二区三区| 国内揄拍国产精品人妻在线| 久久精品夜色国产| 午夜福利在线观看免费完整高清在| 亚洲四区av| 天堂中文最新版在线下载| 女性被躁到高潮视频| 男的添女的下面高潮视频| 少妇 在线观看| 男女边摸边吃奶| 99热全是精品| 婷婷色综合大香蕉| av黄色大香蕉| 日韩av免费高清视频| h视频一区二区三区| 亚洲精品国产成人久久av| 熟女av电影| 嫩草影院新地址| 欧美日本中文国产一区发布| videossex国产| av.在线天堂| 精品一区二区三区视频在线| 国产极品天堂在线| 国产日韩欧美视频二区| av福利片在线| 观看免费一级毛片| 国产精品欧美亚洲77777| 熟女av电影| 免费观看性生交大片5| 女人久久www免费人成看片| 欧美日韩国产mv在线观看视频| 精品熟女少妇av免费看| 97精品久久久久久久久久精品| 国产亚洲5aaaaa淫片| 国产精品久久久久久精品古装| 久久久久精品久久久久真实原创| 一个人免费看片子| 欧美精品国产亚洲| a级片在线免费高清观看视频| 日本爱情动作片www.在线观看| 国产黄片美女视频| 老司机影院成人| 午夜日本视频在线| 中文字幕亚洲精品专区| 最近最新中文字幕免费大全7| 国产淫语在线视频| 人人澡人人妻人| 国产精品.久久久| 国产在视频线精品| 天天操日日干夜夜撸| 韩国av在线不卡| 婷婷色综合大香蕉| 又黄又爽又刺激的免费视频.| 色哟哟·www| 国产伦在线观看视频一区| 国产免费视频播放在线视频| 制服丝袜香蕉在线| 中国美白少妇内射xxxbb| 国产男人的电影天堂91| 中文资源天堂在线| 美女内射精品一级片tv| 天天操日日干夜夜撸| 国产视频内射| 亚洲怡红院男人天堂| 久久久久久久亚洲中文字幕| 少妇的逼水好多| 亚洲情色 制服丝袜| 国产黄片美女视频| 国产淫片久久久久久久久| 蜜臀久久99精品久久宅男| 国产精品蜜桃在线观看| 亚州av有码| 免费看不卡的av| 久久6这里有精品| 国产精品久久久久久久电影| 狂野欧美激情性bbbbbb| 国产一区二区在线观看日韩| 在线观看一区二区三区激情| 国产精品一区二区在线不卡| 人人妻人人澡人人爽人人夜夜| 在现免费观看毛片| 黄片无遮挡物在线观看| videos熟女内射| 99热网站在线观看| 亚洲精品一区蜜桃| 人妻 亚洲 视频| 观看av在线不卡| 麻豆精品久久久久久蜜桃| 久久99一区二区三区| 久久午夜福利片| 九草在线视频观看| 精品久久久久久久久av| 久久久久久久亚洲中文字幕| 亚洲av.av天堂| 日日啪夜夜爽| 菩萨蛮人人尽说江南好唐韦庄| 中文字幕免费在线视频6| 观看av在线不卡| 亚洲国产成人一精品久久久| 久久99蜜桃精品久久| 亚洲美女搞黄在线观看| 少妇的逼好多水| 亚洲在久久综合| 亚洲综合精品二区| 这个男人来自地球电影免费观看 | 久久精品国产亚洲av天美| 少妇丰满av| 五月天丁香电影| 国产一级毛片在线| freevideosex欧美| 在线 av 中文字幕| 国产精品久久久久久久电影| 美女内射精品一级片tv| 国产黄片美女视频| 大片免费播放器 马上看| av福利片在线观看| 丰满少妇做爰视频| 欧美区成人在线视频| 好男人视频免费观看在线| 国产高清不卡午夜福利| 精品人妻熟女av久视频| 视频区图区小说| 中文字幕久久专区| 91在线精品国自产拍蜜月| 欧美xxⅹ黑人| 嘟嘟电影网在线观看| 亚洲精品色激情综合| 色5月婷婷丁香| 一区二区三区乱码不卡18| 日日啪夜夜撸| 久久久久久久久久久丰满| 欧美另类一区| 国产男女超爽视频在线观看| 99热全是精品| 桃花免费在线播放| 亚洲国产精品999| 亚洲va在线va天堂va国产| 欧美亚洲 丝袜 人妻 在线| 十八禁高潮呻吟视频 | 午夜影院在线不卡| 观看美女的网站| 欧美xxⅹ黑人| 欧美 日韩 精品 国产| 欧美精品国产亚洲| 欧美激情极品国产一区二区三区 | 丰满少妇做爰视频| 只有这里有精品99| 少妇精品久久久久久久| 日本av手机在线免费观看| 男女免费视频国产| 国产欧美日韩一区二区三区在线 | 97精品久久久久久久久久精品| 人人妻人人澡人人看| 久久久a久久爽久久v久久| 青青草视频在线视频观看| 午夜福利在线观看免费完整高清在| 免费看av在线观看网站| 又粗又硬又长又爽又黄的视频| 久久久久国产网址| 亚洲欧美一区二区三区国产| 极品人妻少妇av视频| a级片在线免费高清观看视频| 亚洲精品国产av成人精品| 如日韩欧美国产精品一区二区三区 | 乱系列少妇在线播放| av福利片在线| 成人午夜精彩视频在线观看| 好男人视频免费观看在线| 免费高清在线观看视频在线观看| 一个人免费看片子| 啦啦啦中文免费视频观看日本| 国产日韩欧美亚洲二区| 成人18禁高潮啪啪吃奶动态图 | 在线观看免费日韩欧美大片 | 国产有黄有色有爽视频| 成人二区视频| 日本猛色少妇xxxxx猛交久久| 亚洲经典国产精华液单| 人妻系列 视频| 黑人高潮一二区| 久久久久国产网址| 久久精品夜色国产| 成人国产av品久久久| 国产精品一区二区性色av| 欧美日韩亚洲高清精品| 欧美另类一区| 在线天堂最新版资源| 少妇 在线观看| 久久国产乱子免费精品| 国产在线一区二区三区精| 国产av国产精品国产| 一级爰片在线观看| 日本vs欧美在线观看视频 | 高清黄色对白视频在线免费看 | 亚洲精品久久久久久婷婷小说| 啦啦啦啦在线视频资源| 久久精品久久久久久噜噜老黄| 国产日韩欧美亚洲二区| 在线观看三级黄色| 99热这里只有是精品50| 丰满乱子伦码专区| 久久av网站| 卡戴珊不雅视频在线播放| 免费av不卡在线播放| 久久精品久久精品一区二区三区| 女性生殖器流出的白浆| 好男人视频免费观看在线| 亚洲欧美精品自产自拍| 免费黄色在线免费观看| 在线观看免费日韩欧美大片 | 国产乱来视频区| 久久热精品热| 欧美丝袜亚洲另类| 久久久久久久久大av| 观看免费一级毛片| 国产亚洲欧美精品永久| 18+在线观看网站| 丰满迷人的少妇在线观看| 免费大片黄手机在线观看| 国产av国产精品国产| 看十八女毛片水多多多| 我要看黄色一级片免费的| 日韩一本色道免费dvd| 成人二区视频| 人人妻人人澡人人看|