CN105489228A - Rhonchus identification method based on frequency domain image processing - Google Patents

Abstract

The invention relates to a rhonchus identification method based on frequency domain image processing. The method is characterized in that audio signals are converted from time domain into frequency domain through fast Fourier transform, the produced frequency domain images is subjected to binarization processing by the aid of a preset threshold, and specific rhonchus signals are obtained through filtering. The rhonchus identification method has the advantages that rhonchus can be identified fast, the method is high in efficiency, low in calculation amount, low in system load and simple in maintenance, and the blank that currently breath sounds cannot be automatically identified is filled up; by the efficient calculation ability of a computer, the tiny audio frequency which cannot be heard by human ears can be detected, and controversial rhonchus types can be categorized according to specific numerical values.

Description

A kind of rhonchus recognition methods based on frequency domain figure process

Technical field

The present invention relates to signal and image processing field, particularly relate to a kind of rhonchus recognition methods based on frequency domain figure process

Background technology

Rale is a kind of important abnormal breathing sound, by detecting the moment that rale occurs, number, the characteristics such as distribution, doctor can judge the relevant information of doctors and patients' pulmonary disease, and Hospitals at Present medium rales is all by doctor's auscultation discrimination, subjective and be subject to external environment influence.Use robotization identification rale can effectively improve this phenomenon, assist physician improves medical efficiency.

The features such as rhonchus is a branch in rale, has the duration long, and audio frequency distribution is comparatively stable.According to medical facility definition, the rhonchus duration is greater than 250 milliseconds, and distribution frequency is between 60 hertz to 800 hertz.Rhonchus is many to be produced by respiratory tract obst ruction, and doctor judges the diseased region in respiratory tract by rhonchal tone.The same rale duration is long, and frequency stabilization makes it present with horizontal strip shape color lump in spectrogram, and making becomes possibility by the image procossing acquisition rhonchus period.

Summary of the invention

The present invention overcomes above-mentioned weak point, object is to provide a kind of rhonchus recognition methods based on frequency domain figure process, by extracting waveform character in breath sound audio file, use Fourier transform to obtain the frequency domain figure of waveform, filter out unnecessary useless audio frequency by the binaryzation in image procossing thus obtain crucial rhonchus part; The rhonchus in breath sound data can be detected.

The present invention achieves the above object by the following technical programs: a kind of rhonchus recognition methods based on frequency domain figure process, comprises audio frequency pretreatment stage, generates the binaryzation frequency domain figure stage, extracts the feature rhonchus stage;

1) audio frequency pretreatment stage:

1.1) read and storing audio files in voice data;

1.2) noise in filtering voice data, obtains the voice data after denoising;

1.3) carry out data normalization based on the voice data after denoising and remove linear trend;

2) the binaryzation frequency domain figure stage is generated:

2.1) Short Time Fourier Transform treatment step 1.3 is utilized) Wave data that obtains, generate frequency domain figure, and gray processing is carried out to frequency domain figure;

2.2) contrast of frequency domain figure after gray processing is improved;

2.3) each pixel grey scale in frequency domain figure and predetermined threshold value are contrasted, the pixel grey scale lower than threshold value is set to 0, isolates the doubtful rhonchus region in frequency domain figure and non-rhonchus district;

3) the feature rhonchus stage is extracted:

3.1) obtain the feature color lump of the contiguous pixels in doubtful rhonchus region, get rid of the feature color lump of duration lower than Preset Time;

3.2) the feature color lump of gained after analysis and filter one by one, getting rid of the color lump of non-flat elongate in shape, being integrated by last gained color lump, obtaining rhonchus time of occurrence section by analyzing intensity profile.

As preferably, described step 1.1) voice data in audio file is stored with the form of floating-point ordered series of numbers.

As preferably, described step 1.2) utilize external noise in high-pass filtering filtering voice data, utilize the noise that in low-pass filtering filtering voice data, aliasing distortion produces.

As preferably, described step 1.3) carry out the scope of method for utilizing feature Zoom method all values to be converted into [0,1] of data normalization, it is according to raw data data select target scope, and formula is as follows:

$x^{\′}=\frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$

Wherein, x is original value, and x ' is the value after being standardized.

As preferably, described step 1.3) remove the method for linear trend for using FFT preprocess method, after utilizing least-square analysis function to obtain approximate function, from original function, deduct approximate function, obtain the result removing linear trend.

As preferably, described step 2.3) predetermined threshold value determine to obtain according to OTSU adaptive threshold, in OTSU algorithm, use the method for exhaustion to find the value obtaining making maximum between-cluster variance minimum, i.e. the weighted sum of the variance of two classes, formula is as follows:

${\mathrm{\σ}}_w^2\left(t\right)={\mathrm{\ω}}_1\left(t\right){\mathrm{\σ}}_1^2\left(t\right)+{\mathrm{\ω}}_2\left(t\right){\mathrm{\σ}}_2^2\left(t\right)$

Wherein, weights omega _ibe two the distribution probability of class that separates by threshold value t, it is then the variance of each class.

As preferably, described step 2.3) pixel grey scale be 0 region is non-rhonchus district.

As preferably, described step 3.1) Preset Time be 250ms.

As preferably, described step 3.2) to analyze method that intensity profile gets rid of the color lump of non-flat elongate in shape be calculate the distribution situation that feature color lump maps in y-axis, wide if distribute, represent that this time period existed multi-frequency sound superposition phenomenon, got rid of.

Beneficial effect of the present invention is: this method has been filled up at present temporarily without effectively identifying that the field of rhonchus method is blank, in conjunction with Audio Signal Processing and picture signal process related algorithm, makes computing machine automatically can detect rhonchus in breath sound data.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the inventive method;

Fig. 2 is audio frequency pre-service of the present invention and the process schematic generating binaryzation frequency domain figure.

Embodiment

Below in conjunction with specific embodiment, the present invention is described further, but protection scope of the present invention is not limited in this:

Embodiment: as shown in Figure 1, a kind of rhonchus recognition methods based on frequency domain figure process, comprises audio frequency pre-service, and binaryzation frequency domain figure generates, and extracts feature rhonchus; Whole method can realize based on computer mathematics software for calculation MATLAB.

Audio frequency pretreatment operation flow process comprises the following steps:

1) local reading audio file, exists in internal memory by voice data with floating-point array form;

There is related audio function reading wavread to can read wav formatted audio files in MATLAB and corresponding data is converted into matrix ordered series of numbers and be kept at internal memory.

2) some external noises during using high-pass filtering filtering to record;

3) noise that low-pass filtering filtering aliasing distortion produces is used;

Due to the singularity of breath sound file, two-channel is only got any one sound channel and all can.Choose Butterworth filter and bandpass filtering is carried out to sound wave, filter out the sound except 60Hz-2000Hz.Usual clothes friction, health shakes the ambient noise that produces all lower than 60Hz, therefore can effectively extract pure breath sound by filtering, reduces external interference to the impact of testing result.

4) carry out data normalization and remove linear trend;

May be there is the situation that amplitude declines to a great extent in rear section audio frequency after filtering, need with mapminmax function, sound wave matrix to be normalized, promote the volume of the less audio frequency of sound, reduce the volume of the larger audio frequency of sound, ensure that all audio frequency all have identical peak swing.May be there is the situation of displaced central axis in part audio frequency, need to carry out removal linear trend with detrend function, ensures that waveform shakes back and forth along transverse axis.

Normalization is a kind of method be converted into by all values in the scope of [0,1], is also feature convergent-divergent, and it is according to raw data data select target scope, and general formula is as follows:

$x^{\′}=\frac{x-\min \left(x\right)}{\max \left(x\right)-min\left(x\right)}$

X is original value, and x ' is the value after being standardized.

Removing linear trend is a kind of pretreated method of common FFT, deducts to obtain the result of linear trend by using least-square analysis function after obtaining its approximate function from original function.Typical Certain function summary model is linear function model.The simplest linear formula is y=b ₀+ b ₁t, is write as matrix form, for

$\underset{b_0,b_1}{\min }\left|\right|\left(\begin{array}{cc}1& t_1\\ .& .\\ .& .\\ .& .\\ 1& t_n\end{array}\right)\left(\begin{array}{c}{b}_0\\ b_1\end{array}\right)-\left(\begin{array}{c}{y}_1\\ .\\ .\\ .\\ y_n\end{array}\right)|{|}_2=\underset{b}{\min }||Ab-Y|{|}_2.$

Binaryzation frequency domain figure generates:

1) Wave data after using Short Time Fourier Transform process to filter;

2) frequency domain figure generating Short Time Fourier Transform generation carries out gray processing;

MATLAB function spectrogram is used to carry out short time discrete Fourier transform to Audio Matrix and obtain spectrogram after completing pre-service, short time discrete Fourier transform window size is sampling rate/20 (50ms), the large young pathbreaker of window size affects detection degree of accuracy to a great extent, and algorithm predetermined coefficient used is through test of many times gained optimal value.Window size crosses conference makes rhonchus on spectrogram, not present horizontal strip platy character, but diagonally distributes, and has a strong impact on testing result; Otherwise calculated amount then can be caused excessive, drag slow travelling speed.

3) improve generate the contrast of gray-scale map;

Use medium filtering and result and former spectrogram subtracted each other, the method can strengthen the contrast in image between pixel, makes the aberration of zones of different more obvious, promotes test accuracy.Equally in use need after value filtering to be normalized to ensure that signature waveform shock range is enough detected to matrix.

4) gray scale of each pixel and predetermined threshold value are contrasted, the pixel grey scale arranged lower than threshold value is 0; Predetermined threshold value is determined according to OTSU adaptive threshold, and concrete grammar is as follows: in OTSU algorithm, use the method for exhaustion to search out the value that maximum between-cluster variance can be made minimum, namely the weighted sum of the variance of two classes:

${\mathrm{\σ}}_w^2\left(t\right)={\mathrm{\ω}}_1\left(t\right){\mathrm{\σ}}_1^2\left(t\right)+{\mathrm{\ω}}_2\left(t\right){\mathrm{\σ}}_2^2\left(t\right)$

Wherein, weights omega _ibe two the distribution probability of class that separates by threshold value t, and it is then the variance of each class; This algorithm can obtain a value be in [0,1] scope, effectively can isolate the prospect in piece image and background.

In the present embodiment, after completing all pretreatment work, binary conversion treatment is carried out to frequency spectrum gray-scale map, use maximum variance between clusters to obtain the threshold value o of gray-scale map, finding that 0.7*o is optimal threshold through repeatedly testing, effectively can be separated doubtful rhonchus region and non-rhonchus region.So far the image evolution process of step as shown in Figure 2.

Extract feature rhonchus:

1) the feature color lump of contiguous pixels in gray-scale map is obtained;

2) the feature color lump that the eliminating duration is shorter;

The remaining color lump of image after binaryzation is doubtful rhonchus, and cumulative calculation color lump is mapped in the length in x-axis, according to hospital to rhonchal definition length lower than the color lump of 250ms by disallowable.

3) characteristic interval of gained after analysis and filter one by one, gets rid of non-flat by analyzing intensity profile

The color lump of elongate in shape;

In view of the feature of rhonchus frequency stabilization, it will in rectangular flat on spectrogram, the region that each doubtful rhonchus occurs is added up separately, calculate the distribution situation that color lump maps in y-axis, wide if distribute, illustrate that this time period existed multi-frequency sound superposition phenomenon, do not meet rhonchus feature, therefore can be excluded.

4) last gained color lump is integrated, obtain rhonchus time of occurrence section.

In the present embodiment, the sample that experiment test uses sample audio all to use hospital directly to provide, did not do screening and post-production, and 34 groups of audio sample participate in rhonchus test altogether, and totally 26 groups of algorithm recognition results meet diagnosis result, and accuracy rate is 76.5%

Wherein 22 groups is containing rhonchus by diagnosis, and in these 22 groups, by rhonchus algorithm, 17 groups are judged as that accuracy rate is 77.3% containing rhonchus

The know-why being specific embodiments of the invention and using described in above, if the change done according to conception of the present invention, its function produced do not exceed that instructions and accompanying drawing contain yet spiritual time, must protection scope of the present invention be belonged to.

Claims (9)

1. based on a rhonchus recognition methods for frequency domain figure process, it is characterized in that, comprise audio frequency pretreatment stage, generate the binaryzation frequency domain figure stage, extract the feature rhonchus stage;

1) audio frequency pretreatment stage:

1.1) read and storing audio files in voice data;

1.2) noise in filtering voice data, obtains the voice data after denoising;

1.3) carry out data normalization based on the voice data after denoising and remove linear trend;

2) the binaryzation frequency domain figure stage is generated:

2.1) Short Time Fourier Transform treatment step 1.3 is utilized) Wave data that obtains, generate frequency domain figure, and gray processing is carried out to frequency domain figure;

2.2) contrast of frequency domain figure after gray processing is improved;

2.3) each pixel grey scale in frequency domain figure and predetermined threshold value are contrasted, the pixel grey scale lower than threshold value is set to 0, isolates the doubtful rhonchus region in frequency domain figure and non-rhonchus district;

3) the feature rhonchus stage is extracted:

3.1) obtain the feature color lump of the contiguous pixels in doubtful rhonchus region, get rid of the feature color lump of duration lower than Preset Time;

3.2) the feature color lump of gained after analysis and filter one by one, getting rid of the color lump of non-flat elongate in shape, being integrated by last gained color lump, obtaining rhonchus time of occurrence section by analyzing intensity profile.

2. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, is characterized in that: described step 1.1) voice data in audio file is stored with the form of floating-point ordered series of numbers.

3. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, it is characterized in that: described step 1.2) utilize external noise in high-pass filtering filtering voice data, utilize the noise that in low-pass filtering filtering voice data, aliasing distortion produces.

4. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, it is characterized in that: described step 1.3) all values is converted into [0 for utilizing feature Zoom method by the method for carrying out data normalization, 1] scope, it is according to raw data data select target scope, and formula is as follows:

$x^{\′}=\frac{x-min\left(x\right)}{max\left(x\right)-min\left(x\right)}$

Wherein, x is original value, and x ' is the value after being standardized.

5. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, it is characterized in that: described step 1.3) remove the method for linear trend for using FFT preprocess method, after utilizing least-square analysis function to obtain approximate function, from original function, deduct approximate function, obtain the result removing linear trend.

6. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, it is characterized in that: described step 2.3) predetermined threshold value determine to obtain according to OTSU adaptive threshold, in OTSU algorithm, use the method for exhaustion to find the value obtaining making maximum between-cluster variance minimum, the i.e. weighted sum of the variance of two classes, formula is as follows:

${\mathrm{\σ}}_w^2\left(t\right)={\mathrm{\ω}}_1\left(t\right){\mathrm{\σ}}_1^2\left(t\right)+{\mathrm{\ω}}_2\left(t\right){\mathrm{\σ}}_2^2\left(t\right)$

Wherein, weights omega _ibe two the distribution probability of class that separates by threshold value t, it is then the variance of each class.

7. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, is characterized in that: described step 2.3) pixel grey scale be 0 region is non-rhonchus district.

8. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, is characterized in that: described step 3.1) Preset Time be 250ms.

9. a kind of rhonchus recognition methods based on frequency domain figure process according to claim 1, it is characterized in that: described step 3.2) to analyze method that intensity profile gets rid of the color lump of non-flat elongate in shape be calculate the distribution situation that feature color lump maps in y-axis, wide if distribute, represent that this time period existed multi-frequency sound superposition phenomenon, got rid of.