CN110542658A - A classification method of tobacco non-smoke substances based on hyperspectral imaging technology - Google Patents

Abstract

A method for classifying tobacco non-smoke substances based on hyperspectral imaging technology, which is characterized in that: using short-wave imaging hyperspectral technology to first classify tobacco leaves and sundries, establish a spectral library including different substances, and then collect image data from samples to be tested , use the reference spectrum in the spectral library to match the measured sample, and judge it, and then complete the effective classification and identification of tobacco leaves and sundries. The invention combines spectral technology and two-dimensional image imaging technology. Compared with the prior art, it has the following significant improvements: 1. The present invention uses the short-wave imaging hyperspectral technology to classify tobacco leaves and sundries, and establishes a spectral library containing different substances. 2. The experimental process of the present invention does not use toxic and harmful chemicals, which is simple, fast, non-destructive to the sample, and non-polluting to the environment. 3. The present invention has the advantages of simple operation, rapidity, accuracy, low cost and high efficiency.

Description

A classification method of tobacco non-smoke substances based on hyperspectral imaging technology

technical field

The invention belongs to the field of hyperspectral agricultural production and application, in particular to a method for classifying tobacco non-smoke substances based on hyperspectral imaging technology.

Background technique

Hyperspectral imaging is a comprehensive technology that integrates detector technology, precision optical machinery, weak signal detection, computer technology, and information processing technology. Hyperspectral imaging simultaneously detects the two-dimensional geometric space and one-dimensional spectral information of the target, and obtains continuous and narrow-band image data with high spectral resolution. It is a multi-dimensional information acquisition technology that combines imaging technology with spectral technology. Therefore, the use of hyperspectral imaging technology can accurately reflect the object. In the tobacco industry, the quality and purity of tobacco raw materials are directly related to the quality of cigarette products. If tobacco leaves are mixed with non-tobacco sundries during production, acquisition, transportation and processing (Class I sundries: metal, feather, plastic, etc.; Class II sundries: paper, stone, hemp rope, glass, etc.; Class III sundries: Non-tobacco leaves, seeds, bamboo sticks, etc.) will cause great hidden dangers to the quality of cigarette products processed by the cigarette industry.

At present, the methods of impurity removal used in cigarette production and processing lines mainly include: wind impurity removal, photoelectric impurity removal, magnetic impurity removal and manual selection of impurity removal. Attribute differences are identified and eliminated in a targeted manner, such as differences in specific gravity, color differences and magnetic differences. Therefore, a method of removing impurities can only identify and remove a certain type of debris. Manual removal of impurities can identify most of the non-smokers, but the human eyes are easily fatigued and work efficiency is low.

The unique chemical composition and physical detection characteristics of tobacco leaves lead to unique diagnostic characteristic absorption bands, which have relatively stable wavelength positions and unique waveforms. From the chemical point of view of tobacco leaves, the main chemical components in tobacco leaves are carbohydrates, nitrogen compounds, organic acids and minerals. Among them, the absorption characteristics of carbon water are at 1000-2500nm, and the absorption characteristics of nitrogen compounds are at 1500nm-1750nm. As important chemical components of tobacco leaves, the two are not reflected in the visible light band (400-1000nm). That is, the spectrum of tobacco leaves will form strong absorption peaks near carbon, water and nitrogen compounds, and these absorption peaks can be used as an important basis for distinguishing tobacco leaves. Using spectral information technology, non-tobacco leaf debris and tobacco leaves form different specific spectra in the same optical environment, identify different substances, and complete the classification of tobacco leaves and debris.

Chinese Patent (200910059486.5) discloses a method for rapid determination of the proportion of tobacco cut tobacco based on a near-infrared spectrometer. Compared with the method proposed in this patent, there are two main differences between the two: First, the scope of application of the two is different. The Chinese patent (200910059486.5) is mainly used for the determination of the proportion of shredded tobacco, and the method in this paper is mainly used for the identification of impurities in tobacco leaves; Hyperspectral imager) This method is more convenient and quick to operate, without the need for sample crushing, screening and other operations, and can be marked online.

Chinese patent (201410491816.9) mainly discloses a method for identifying components of cut tobacco based on spectral imaging technology. Compared with the method proposed in this patent, there are the following differences. 1. The instrument used in the method proposed in this paper is a hyperspectral imager with a resolution of 12 nm and a higher identification accuracy. 2. The spectroscopic camera used in this patent is composed of a band-pass filter and a CCD camera with a fixed-focus lens. The customized wavelength band of the band-pass filter is only for imaging and image processing analysis of cut tobacco, and cannot be analyzed for other substances. and identification. 3. The method proposed in this patent is only used to determine the components of cut tobacco (cut leaf, cut stem, reconstituted tobacco leaf, and expanded cut leaf), and the method proposed in this paper is used for the identification of tobacco leaves and non-tobacco substances, and the two are inconsistent in function. .

In summary, the use of spectroscopic techniques to determine the content of cut tobacco and the quantitative determination of chemical substances in tobacco by hyperspectral have certain research and applications in the industry. However, using spectral information technology, according to the formation of different specific spectra of non-tobacco leaf debris and tobacco leaves in the same optical environment, different substances are identified, and the classification of tobacco leaves and non-tobacco substances is still blank.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for classifying tobacco non-smoke substances based on hyperspectral imaging technology based on the above state of the art, which can quickly and accurately distinguish tobacco leaves and sundries, and prevent non-smoke sundries from entering the finished tobacco products. .

The purpose of this invention is to realize through the following technical solutions:

A method for classifying and identifying tobacco leaf debris based on hyperspectral imaging is to use short-wave hyperspectral imaging technology to first classify tobacco leaves and debris, establish a spectral library including different substances, and then collect hyperspectral image data of the sample to be tested, and use The reference spectrum in the spectral library matches the sample to be tested and judges it, thereby completing the effective classification and identification of tobacco leaves and sundries.

Specific steps are as follows:

1) Sample collection: Obtain samples of Class I, Class II and Class III sundries including pure tobacco leaves;

The surface of the sample is kept dry and clean with no other attachments.

The first-class sundries include metals, feathers, and plastics; the second-class sundries include paper, stone, hemp rope, and glass; and the third-class sundries include non-tobacco leaves, seeds, and bamboo sticks.

2) Sample preparation and hyperspectral imaging with black and white frame correction;

Using a tungsten halogen lamp as the illumination light source, hyperspectral image acquisition was performed on the obtained tobacco leaf and debris samples to obtain hyperspectral images of the samples. To reduce the effect of noise, black and white frame correction is performed on the hyperspectral image. The black and white frame correction formula is as follows:

In the formula: R-corrected hyperspectral image; I-original hyperspectral image; B-completely black image collected by turning off the camera lens; W-completely white image obtained by scanning the white correction plate.

3) Hyperspectral image preprocessing and acquisition of characteristic images;

In order to improve the signal-to-noise ratio of the data, the hyperspectral image data is preprocessed, and the preprocessing methods include but are not limited to:

By using the Savitzky-Golay convolution smoothing filtering algorithm, the baseline drift and tilt are removed, noise is removed, and the smoothness of the spectral curve is improved;

Then, the scattering effect on the surface of the object is reduced by multivariate scattering correction (MSC), and the spectral absorption information between the same substances is enhanced.

4) Extract the spectral information of the sample and establish a spectral library file;

The specific process includes but is not limited to:

The tobacco leaf and foreign matter samples include: pure tobacco leaf samples, first-class foreign matter samples, second-class foreign matter samples, and third-class foreign matter samples;

The tobacco leaves and the regions of interest (ROI) of the first, second and third types of debris were selected respectively, and the spectral features of the selected sample regions of interest were extracted, and the average spectrum was obtained;

A spectral information library file is established, and the average spectrum obtained in the sample ROI area is imported into the library file for saving.

5) The realization of tobacco leaf debris classification:

Principal component analysis (PCA) was used to reduce the dimension of the hyperspectral image of the collected sample, and then the target spectrum was matched by the spectral angle matching method (SAM). Label different samples.

The specific process is as follows: the tobacco leaves mixed with one, two, and three types of impurities are imaged with a short-wave hyperspectral imager, and the collected data is preprocessed and compared with the spectral information recorded in the spectral library. Dimensionality reduction and feature recognition of tobacco leaves and sundries on hyperspectral images by using spectral angle matching (PCA) and spectral angle matching (SAM) algorithms, including:

First, scan and obtain short-wave hyperspectral imaging information of the sample;

Secondly, adopt the principal component analysis method to reduce the dimension of the collected samples;

Finally, through the calculation of the spectral angle matching algorithm, it is judged whether the tobacco leaf sample is mixed with foreign matter according to the spectral characteristic vector, and different samples are marked.

The preprocessing method in step 3) can also adopt mean centering (mean centering), standardization (autoscaling), normalization (normalization), standard normal variable transformation (SNV), derivative, smoothing denoising algorithm, wavelet transform. a kind of

The invention provides a tobacco leaf debris classification method based on short-wave imaging hyperspectral technology, which combines spectral technology and two-dimensional image imaging technology. Compared with the existing technology, it has the following significant progress:

1. The present invention uses short-wave imaging hyperspectral technology to classify tobacco leaves and sundries, and establishes a spectral library containing different substances.

2. The experimental process of the present invention does not use toxic and harmful chemicals, which is simple, fast, non-destructive to the sample, and non-polluting to the environment.

3. The present invention has the advantages of simple operation, rapidity, accuracy, low cost and high efficiency.

Description of drawings

1 is a flowchart of a method for identifying and classifying sundries in tobacco leaves based on short-wave hyperspectral imaging technology provided by the present invention;

Figure 2 is the spectrum of the mixed impurities of tobacco leaves.

Detailed ways

The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

Experimental equipment and parameters

The SWIR hyperspectral imager produced by GaiaSorter-N25E of Sichuan Shuangli Hepu Technology Co., Ltd. has a wavelength range of 1000-2500nm, a spectral resolution of 12nm, an image resolution of 384*288 pixels, and a spectrometer frame number of 400. The exposure time is 20ms; the experimental platform uses a GaiaSorter-Dual type dark box system, which includes a halogen tungsten light source, an electronic control platform, and the like. In order to avoid the influence of external stray light sources, the hyperspectral image acquisition process is carried out in a dark box system.

Step 1: Collection of samples

In order to make the established spectral database more widely applicable, this example selects the first and second types of sundries, such as cigarettes from a certain China Tobacco flue-curing machine in 2018, and feathers, cloth strips, and plastic ropes provided by a certain China Tobacco.

Step 2: Sample Preparation and Hyperspectral Imaging with Black and White Frame Correction

Select pure tobacco leaves, cloth strips, feathers, and plastic ropes with no attachments on the surface after dust removal, lay them flat on the black backboard, and mark their locations. Using the SWIR hyperspectral imager produced by Shuangli Hepu Company, firstly close the lens cover and collect the full black calibration image with zero reflectivity. Then open the lens cover, scan a standard white board, acquire an all-white calibration image with an emissivity of 99.9%, and then perform hyperspectral imaging on tobacco leaves and debris samples. The SpecView software is used to collect the image, and black and white frame correction is performed on the collected hyperspectral image of the sample according to the collected all-black and all-white calibration images to reduce the noise caused by external stray light, and finally it is used as the original image. Spectral data format for storage.

The black and white frame correction formula is as follows:

In the formula: R-corrected hyperspectral image; I-original hyperspectral image; B-all-black calibration image collected by turning off the camera lens; W-all-white calibration image obtained by scanning the white calibration plate.

Step 3: Hyperspectral image preprocessing and acquisition of characteristic images

In order to remove baseline drift and tilt, remove the influence of noise, and improve the smoothness of the spectral curve, the Savitzky-Golay convolution smoothing filtering algorithm is used.

Savitzky-Golay filtering can improve the smoothness of the spectral curve and reduce the interference of noise. The key to its smooth convolution lies in the solution of the matrix operator. The mean square error (MSE) is introduced to select the appropriate window width n and the number of filter kernel center points m. The smaller the MSE value, the smaller the noise. Let the width of the filter window be n=2m+1, and each measurement point is x=(-m,-m+1...0...m-1,m), and use k-1 polynomial to fit the data points in the window. combine:

y=a ₀ +a ₁ x+a ₂ x ² +…+a _k-1 x ^k-1

There are n above equations, which form a system of k-element linear equations. In order to make the system of equations have a solution, let n>k, and solve the fitting parameter A by the least square method, which can be obtained:

It is represented by a matrix as:

Y _(2m+1)·1 =X _(2m+1)·k ·A _k·1 +E _(2m+1)·1

The least squares solution of A for:

Model predicted or filtered value of Y for:

The spectral characteristics of the collected samples can be corrected by using multiple scattering correction (MSC) to improve the spectral signal-to-noise ratio.

Calculate the average spectrum of the spectrum to be corrected:

Unary linear regression:

Multivariate Scatter Correction:

In the formula, A is the calibration spectrum data matrix, n is the number of calibration samples, and p is the number of wavelength points when the spectrum is collected; - Average spectral vector.

Step 4: Extract the spectral information of the sample and create a spectral library file

The hyperspectral images were collected for the first, second and third types of debris including pure tobacco leaf samples, and the region of interest was selected between the different samples collected, the average spectral information between the samples was extracted, and the spectral library file was established.

Step 5: Realization of Tobacco Leaf Debris Classification

Principal component analysis (PCA) was used to reduce the dimension of the collected sample hyperspectral images, and then the spectral angle matching (SAM) algorithm was used to match the sample spectra, so as to achieve the purpose of classifying tobacco leaf debris.

Principal Component Analysis (PCA) selects eigenvectors to achieve dimensionality reduction of multi-index vectors. Its dimensionality reduction model is as follows:

y ₁ =l ₁₁ x ₁ +l ₁₂ x ₂ +...+l _1n x _n

y ₂ =l ₂₁ x ₁ +l ₂₂ x ₂ +...+l _2n x _n

y _m =l _m1 x ₁ +l _m2 x ₂ +…+l _mn x _n

In the formula, x is the n-dimensional multi-index vector, and y is the m-dimensional principal component vector obtained after processing.

The principal component analysis eigenvalues and variance contribution rates in the examples are shown in Table 1 below.

Table 1 Principal component analysis eigenvalues and variance contribution rate

The spectral angle matching method uses the reference spectrum in the spectral library to match the unknown sample, and judges it by setting the narrow-sense spectral angle threshold. Its formula is:

where T is the standard principal component score vector, and R is the reference principal component score vector (both T and R are non-zero vectors).

Further, the spectral information of the hyperspectral image of the line push-broom imaging is compared with the spectral library established by the supervised classification to obtain the identification result.

The classification and identification method based on hyperspectral imaging provided by this embodiment at least includes the following beneficial effects:

1. Use Savitzky-Golay filtering to improve the smoothness of the spectral curve and reduce noise interference; use multiple scattering correction to improve the signal-to-noise ratio of the spectral signal; both provide the basis for subsequent image processing work

2. Obtain spectral information of different substances through hyperspectral image acquisition. Create a spectral library file to save sample spectral information.

3. The method of principal component analysis is used to reduce the dimension, reduce the amount of image calculation, eliminate the influence of non-important features, and greatly reduce the calculation time. Through the spectral angle matching algorithm, the reference spectrum in the spectral library is used to match the unknown sample, and the narrow-sense spectral angle threshold is used to judge it, so as to complete the effective classification and identification of tobacco leaves and sundries, so as to obtain a more accurate classification effect.

4. Compared with the traditional manual impurity removal, the work efficiency and detection speed are improved. At the same time, the use of hyperspectral imaging technology can obtain finer composition information and improve the accuracy of resolution.

The present invention has been described in detail above in conjunction with the embodiments, rather than limiting the embodiments of the present invention. Within the scope of knowledge possessed by those of ordinary skill in the art, it can also be made without departing from the principles and purposes of the present invention. Various changes, modifications, substitutions and deformations. The scope of the invention is defined by the claims and their equivalents.

Claims (10)

1. A tobacco non-smoke substance classification method based on a hyperspectral imaging technology is characterized by comprising the following steps: tobacco leaves and impurities are classified by utilizing a short-wave hyperspectral imaging technology, a spectrum library comprising different substances is established, hyperspectral image data are collected for a sample to be detected, a reference spectrum in the spectrum library is used for matching the sample to be detected and judging the sample to be detected, and then effective classification and identification of the tobacco leaves and the impurities are completed.

2. The hyperspectral imaging technology-based tobacco non-smoke substance classification method according to claim 1, characterized in that: the method comprises the following specific steps:

1) Collecting samples: obtaining samples of first-class sundries, second-class sundries and third-class sundries including pure tobacco leaves;

2) Sample preparation and hyperspectral imaging are carried out, and black and white frame correction is carried out;

3) preprocessing a hyperspectral image and acquiring a characteristic image;

4) Extracting sample spectrum information and establishing a spectrum library file;

5) the tobacco leaf sundries are classified:

performing dimensionality reduction on a hyperspectral image of an acquired sample by adopting a Principal Component Analysis (PCA), matching a target spectrum by utilizing a spectral angle matching method (SAM), judging whether impurities are mixed in the tobacco sample or not according to a spectral feature vector, and marking different samples.

3. The tobacco non-smoke substance classification method based on the hyperspectral imaging technology according to claim 2, characterized in that: and 2) using a halogen tungsten lamp as an illumination light source, and performing hyperspectral image acquisition on the obtained tobacco leaf and sundry samples to obtain hyperspectral images of the samples. In order to reduce the noise influence, the hyperspectral image is corrected by a black and white frame correction formula as follows:

In the formula: r-a corrected hyperspectral image; i-an original hyperspectral image; b, closing the all-black image collected by the camera lens; w-scanning the white correction plate to obtain a full white image.

4. The tobacco non-smoke substance classification method based on the hyperspectral imaging technology according to claim 2, characterized in that: in step 3), in order to improve the signal-to-noise ratio of the data, the hyperspectral image data is preprocessed, and the preprocessing method includes but is not limited to:

By using a Savitzky-Golay convolution smoothing filtering algorithm, baseline drift and inclination are removed, noise is removed, and smoothness of a spectral curve is improved;

And then reducing the scattering effect on the surface of the object by Multivariate Scattering Correction (MSC) to enhance the spectral absorption information among the same substances.

5. the tobacco non-smoke substance classification method based on the hyperspectral imaging technology according to claim 2, characterized in that: in step 4), the specific processes include but are not limited to:

The tobacco leaf and sundries sample comprises: a pure tobacco leaf sample, a first class sundry sample, a second class sundry sample and a third class sundry sample;

Respectively selecting tobacco leaves and regions of interest (ROI) of the first, second and third sundries, extracting spectral characteristics at the selected region of interest of the sample, and obtaining an average spectrum;

and establishing a spectrum information library file, and importing the average spectrum obtained from the ROI area of the sample into the library file for storage.

6. The tobacco non-smoke substance classification method based on the hyperspectral imaging technology according to claim 2, characterized in that: the specific process in step 5) is as follows: the tobacco leaves mixed with the first, second and third impurities are subjected to image acquisition by a short wave hyperspectral imager, the acquired data are preprocessed and then compared with spectral information recorded in a spectral library, and the hyperspectral images are subjected to dimensionality reduction and characteristic identification of the tobacco leaves and the impurities by a Principal Component Analysis (PCA) and a Spectral Angle Matching (SAM) algorithm, wherein the method comprises the following steps:

Firstly, scanning and obtaining short-wave hyperspectral imaging information of the sample;

secondly, performing dimensionality reduction on the collected sample by adopting a principal component analysis method;

And finally, calculating by a spectral angle matching algorithm, judging whether impurities are mixed in the tobacco leaf sample according to the spectral feature vector, and marking different samples.

7. The hyperspectral imaging technology-based tobacco non-smoke substance classification method according to claim 4, characterized in that: the preprocessing method in step 3 may further adopt one of mean centering (mean centering), normalization (normalization), standard normal variable transformation (SNV), derivative, smooth denoising algorithm, and wavelet transformation.

8. The tobacco non-smoke substance classification method based on hyperspectral imaging technology according to claim 1 or 2, characterized in that: the sample surface remained dry and clean with no other attachments.

9. the tobacco non-smoke substance classification method based on hyperspectral imaging technology according to claim 1 or 2, characterized in that: the hyperspectral imager is used for hyperspectral imaging, the wavelength range is 1000-2500nm, the spectral resolution is 12nm, the image resolution is 384 × 288 pixels, and the spectrometer frame number is 400.

10. A tobacco non-smoke substance classification method based on hyperspectral imaging technology according to claim 2 or 5, characterized in that: the sundries comprise metal, feather and plastic; the second class of sundries comprises paper, stones, hemp ropes and glass; the three kinds of impurities comprise non-tobacco leaves, seeds and bamboo sticks.