Tobacco online near infrared spectrum detection device and method based on linear optical filter

Abstract

The invention discloses a tobacco online near infrared spectrum detection device and method based on a linear optical filter, and relates to the technical field of spectrum processing. Near-infrared light is emitted to a target area, a plurality of near-infrared bands are obtained through a linear optical filter, and reflected light of each near-infrared band is collected to obtain an original near-infrared band set; preprocessing the target original near-infrared band to obtain a target near-infrared band; performing feature extraction on the target near-infrared band to obtain target sub-features; and constructing a target feature vector according to all the target sub-features, and searching a preset database according to the target feature vector to obtain a detection result. Multiple near-infrared bands and frequency-combining and frequency-doubling absorption information of substance molecules corresponding to different bands are obtained through the linear optical filter, and then the different near-infrared bands are preprocessed and then subjected to feature extraction, so that prediction of different chemical substance components in tobacco is improved, and the online detection efficiency of the tobacco is improved.

Description

Technical Field

[0001] This invention belongs to the field of spectral processing technology, specifically relating to an online near-infrared spectral detection device and method for tobacco based on a linear filter. Background Technology

[0002] During cigarette processing, fluctuations in the chemical composition of tobacco leaves directly impact the stability of cigarette product quality and style characteristics. Rapid and accurate detection of tobacco leaf chemical components provides a reliable basis for raw material input and equipment parameter adjustments, helping to improve process control in cigarette processing enterprises and achieve homogenization of cigarette product quality. Real-time online monitoring and detection of tobacco chemical components are essential in all stages of tobacco processing to ensure stable production and uniform product quality. Online detection, as an effective way to achieve automated and intelligent production, is widely used in the tobacco industry. Currently, the tobacco industry mainly uses near-infrared spectroscopy for online chemical component detection. Specifically, a near-infrared probe is mounted on a material conveyor belt, collecting near-infrared information as the material flows. Near-infrared spectroscopy is an indirect analysis technique. It establishes a calibration model between the sample's near-infrared spectrum and a reference value and compares them to complete qualitative or quantitative analysis of the sample. However, the online near-infrared spectrometers currently used in the tobacco industry mostly employ grating or Fourier transform spectrometers, which suffer from poor shock resistance, poor structural stability, and long detection cycles.

[0003] Patent CN116879221A discloses a method for analyzing the elemental content of tobacco based on near-infrared spectroscopy. The method involves: a) collecting near-infrared spectra of various tobacco samples; b) detecting the C, O, and N elemental contents of each tobacco sample; c) constructing models, dividing 49 tobacco samples into training and test sets; using chemometric algorithms to establish prediction models for the C, O, and N elemental contents of tobacco based on near-infrared spectroscopy for the training set; d) validating the models, inputting the near-infrared spectra of 12 test set samples into the three prediction models obtained in step c, and comparing the predicted C, O, and N elemental contents with those measured in step b to verify and evaluate the model's accuracy; and e) applying the models, detecting near-infrared spectra of tobacco samples with unknown elemental analysis results using the method in step a, and substituting the results into the model to obtain the predicted C, O, and N elemental contents for the corresponding samples. However, this method still suffers from inaccurate subsequent chemical analysis when near-infrared spectral acquisition is interfered with, resulting in low efficiency in online tobacco detection. Summary of the Invention

[0004] The purpose of this invention is to solve the problem that interference with near-infrared spectral acquisition still leads to inaccurate subsequent chemical analysis and low efficiency in online tobacco detection. Therefore, this invention proposes an online near-infrared spectral detection device and method for tobacco based on a linear filter.

[0005] In a first aspect of this invention, a method for online near-infrared spectroscopy detection of tobacco based on a linear filter is first proposed, the method comprising: Near-infrared light is emitted toward the target area and passed through a linear filter to obtain multiple near-infrared bands. The reflected light from each near-infrared band is collected to obtain the original near-infrared band set. The target's original near-infrared band is obtained by preprocessing the original near-infrared band; the original near-infrared band is any one of the original near-infrared bands in the set of original near-infrared bands. Feature extraction is performed on the near-infrared band of the target to obtain target sub-features; A target feature vector is constructed based on all target sub-features, and the detection result is obtained by searching a preset database based on the target feature vector.

[0006] Optionally, preprocessing the original near-infrared band of the target to obtain the target's near-infrared band includes: The target line equation is obtained by performing a linear approximation fit on the original near-infrared band, and the baseline correction value of the original near-infrared band is determined based on the linear equation. The noise spectrum is obtained by extracting noise from the near-infrared band based on the baseline correction value. The target near-infrared band is obtained by correcting the original near-infrared band based on the noise spectrum.

[0007] Optionally, feature extraction of the target in the near-infrared band to obtain target sub-features includes: Substituting the target's near-infrared band into the preprocessing layer yields the first preprocessing feature; The first preprocessed feature is successively substituted into two residual blocks to obtain the first residual feature, and the first residual feature is substituted into the improved residual block to obtain the second residual feature; Substituting the second residual feature into the residual block yields the third residual feature, and substituting the third residual feature into the improved residual block yields the fourth residual feature; The target sub-feature is obtained by performing convolutional pooling on the fourth residual feature.

[0008] Optionally, substituting the target near-infrared band into the preprocessing layer to obtain the first preprocessing feature includes: The first convolutional feature is obtained by performing a 1×5 convolution operation on the near-infrared band of the target. The first convolutional features are normalized and then fed into the max pooling layer to obtain the first preprocessed features.

[0009] Optionally, the working principle of the improved residual block includes: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the second convolutional features; After batch processing the second convolutional features, the ReLU activation function is input to obtain the activated features; Perform a 3×3 convolution operation on the activation features to obtain the third convolutional feature; The input features are subjected to a 1×3 convolution operation to obtain the fourth convolution feature. The third convolution feature and the fourth convolution feature are then added element-wise to obtain the output feature of the preprocessing layer.

[0010] In a second aspect of the invention, an online near-infrared spectroscopy detection device for tobacco based on a linear filter is provided, comprising: The original near-infrared band acquisition module is used to emit near-infrared light into the target area and obtain multiple near-infrared bands through a linear filter. The reflected light of each near-infrared band is collected to obtain the original near-infrared band set. The preprocessing module is used to preprocess the original near-infrared band of the target to obtain the target near-infrared band; the original near-infrared band of the target is any one of the original near-infrared bands in the set of original near-infrared bands; The feature extraction module is used to extract features from the near-infrared band of the target to obtain target sub-features; The detection result determination module is used to construct a target feature vector based on all target sub-features, and to search a preset database based on the target feature vector to obtain the detection result.

[0011] Optionally, the preprocessing module includes: The baseline correction value determination module is used to perform a linear approximation fit on the original near-infrared band to obtain the target straight line equation, and determine the baseline correction value of the original near-infrared band based on the straight line equation. The noise spectrum extraction module is used to extract noise from the near-infrared band based on the baseline correction value to obtain a noise spectrum; The near-infrared correction module is used to correct the original near-infrared band according to the noise spectrum to obtain the target near-infrared band.

[0012] Optionally, the feature extraction module includes: The first preprocessing feature determination module is used to substitute the target near-infrared band into the preprocessing layer to obtain the first preprocessing feature; The second residual feature generation module is used to continuously substitute the first preprocessed feature into two residual blocks to obtain the first residual feature, and substitute the first residual feature into the improved residual block to obtain the second residual feature. The fourth residual feature generation module is used to substitute the second residual feature into the residual block to obtain the third residual feature, and substitute the third residual feature into the improved residual block to obtain the fourth residual feature; The target sub-feature generation module is used to perform convolutional pooling operations on the fourth residual feature to obtain the target sub-feature.

[0013] Optionally, the first preprocessing feature determination module includes: The first convolutional feature determination module is used to perform a 1×5 convolution operation on the near-infrared band of the target to obtain the first convolutional feature. The first preprocessing feature generation module is used to normalize the first convolutional features and then feed them into the max pooling layer to obtain the first preprocessing features.

[0014] Optionally, the improved residual block includes: The feature acquisition module is used to acquire input features and perform a 1×1 convolution operation on the input features to obtain a second convolution feature. The activation module is used to batch process the second convolutional features and then input them into the ReLU activation function to obtain activated features; The convolution module is used to perform a 3×3 convolution operation on the activation features to obtain a third convolution feature; The output module is used to perform a 1×3 convolution operation on the input features to obtain a fourth convolution feature, and to add the third convolution feature and the fourth convolution feature element by element to obtain the output feature of the preprocessing layer.

[0015] The beneficial effects of this invention are: This invention proposes an online near-infrared spectroscopy detection method for tobacco based on a linear filter. Near-infrared light is emitted towards the target area and passed through a linear filter to obtain multiple near-infrared bands. The reflected light from each near-infrared band is collected to obtain a raw near-infrared band set. The raw near-infrared bands are preprocessed to obtain the target near-infrared band. The target raw near-infrared band is any one of the raw near-infrared bands in the raw near-infrared band set. Feature extraction is performed on the target near-infrared band to obtain target sub-features. A target feature vector is constructed based on all target sub-features, and the detection result is obtained by searching a preset database based on the target feature vector. By acquiring multiple near-infrared bands through a linear filter, the combination and overtone absorption information of the corresponding molecules in different bands is obtained. Furthermore, feature extraction is performed after preprocessing different near-infrared bands, thereby improving the prediction of different chemical components in tobacco and increasing the efficiency of online tobacco detection. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 A flowchart of an online near-infrared spectroscopy detection method for tobacco based on a linear filter, provided in an embodiment of the present invention; Figure 2 A flowchart of another method for online near-infrared spectroscopy detection of tobacco based on a linear filter provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an online near-infrared spectroscopy detection device for tobacco based on a linear filter, provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides an online near-infrared spectroscopy detection method for tobacco based on a linear filter. See also... Figure 1 , Figure 1 This is a flowchart illustrating an online near-infrared spectroscopy detection method for tobacco based on a linear filter, provided as an embodiment of the present invention. The method includes the following steps: S101 emits near-infrared light toward the target area and obtains multiple near-infrared bands through a linear filter. The reflected light of each near-infrared band is collected to obtain the original near-infrared band set. S102, preprocess the original near-infrared band of the target to obtain the target's near-infrared band; S103, extract target sub-features by performing feature extraction on the near-infrared band of the target; S104: Construct a target feature vector based on all target sub-features, and obtain the detection result by searching a preset database based on the target feature vector.

[0021] Among them, the target's original near-infrared band is any one of the original near-infrared bands in the set of original near-infrared bands.

[0022] The present invention provides an online near-infrared spectroscopy detection method for tobacco based on a linear filter. This method acquires multiple near-infrared bands using a linear filter, and obtains the combination and overtone absorption information of molecules corresponding to different bands. After preprocessing different near-infrared bands, feature extraction is performed, thereby improving the prediction of different chemical components in tobacco and increasing the efficiency of online tobacco detection.

[0023] In one implementation, the combination and overtone absorption information of molecules in different near-infrared bands are obtained through a linear filter. Multi-band data can more comprehensively characterize the composition of substances and avoid missing information in a single band. The entire process from light emission to obtaining detection results can be completed in seconds, which is suitable for real-time quality monitoring on the production line and significantly shortens the detection cycle compared with traditional chemical analysis (such as chromatography and mass spectrometry).

[0024] In one implementation, preprocessing the original band can eliminate interference such as instrument noise and differences in sample surface reflection, improve the signal-to-noise ratio, make subsequent feature extraction more accurate, and significantly improve the repeatability and reliability of the detection results.

[0025] In one implementation, multiple near-infrared bands are obtained through a linear filter. The specific properties of the linear filter are determined by technicians. The linear filter can divide near-infrared light into different near-infrared bands, thereby enabling the analysis of different chemical substances in tobacco.

[0026] In one implementation, by extracting the features of the near-infrared band of each target, key information can be screened from high-dimensional spectral data, reducing the data dimensionality; then, by constructing feature vectors and matching them with a preset database (establishing the chemical substance content corresponding to the target sub-features of different near-infrared spectra based on historical data), quantitative analysis can be achieved, improving the accuracy of chemical substance content detection.

[0027] In one embodiment, preprocessing the original near-infrared band of the target to obtain the target's near-infrared band includes: S1021, The target line equation is obtained by performing a linear approximation fitting on the original near-infrared band, and the baseline correction value of the original near-infrared band is determined based on the line equation. S1022, noise spectrum is obtained by noise extraction in the near-infrared band based on the baseline correction value; S1023, the target near-infrared band is obtained by correcting the original near-infrared band according to the noise spectrum.

[0028] In one implementation, the original near-infrared band often experiences baseline drift due to sample condition, instrument stability, or environmental factors, resulting in distortion of absorption peak position and intensity. By approximating the target straight line equation through a straight line, the spectral baseline can be restored to an ideal straight line. If the original spectral baseline is inclined, the deviation value (baseline correction value) between each wavelength point and the straight line can be calculated after straight line fitting, thereby eliminating the influence of baseline shift on the real signal.

[0029] In one implementation, the noise distribution characteristics can be quantified by calculating the noise spectrum of each band based on the baseline correction value; the correction process is to subtract the noise spectrum from the original spectrum.

[0030] In one implementation, a sliding window (determined by a technician) is set on the baseline correction value, and the standard deviation of the baseline correction value in each window is calculated as an estimate of the noise intensity in that area; the local standard deviation is proportionally distributed to each wavelength point in the near-infrared band to generate a noise spectrum.

[0031] In one embodiment, feature extraction of the target in the near-infrared band to obtain target sub-features includes: Substituting the target's near-infrared band into the preprocessing layer yields the first preprocessing feature; The first preprocessed feature is successively substituted into two residual blocks to obtain the first residual feature, and the first residual feature is substituted into the improved residual block to obtain the second residual feature; Substituting the second residual feature into the residual block yields the third residual feature, and substituting the third residual feature into the improved residual block yields the fourth residual feature. The target sub-feature is obtained by performing convolutional pooling on the fourth residual feature.

[0032] In one implementation, preprocessing the target's near-infrared band improves the quality of subsequent feature inputs; the residual block directly transmits shallow features to deep layers through skip connections, allowing gradients to bypass multiple nonlinear transformations and directly backpropagate to extract features from different receptive fields.

[0033] In one implementation, the convolution pooling operation on the fourth residual feature is performed by performing a 1×1 convolution operation on the fourth residual feature.

[0034] In one embodiment, substituting the target's near-infrared band into the preprocessing layer to obtain the first preprocessing feature includes: The first convolutional feature is obtained by performing a 1×5 convolution operation on the near-infrared band of the target. The first convolutional features are normalized and then fed into the max pooling layer to obtain the first preprocessed features.

[0035] In one implementation, in the near-infrared spectrum, the characteristic absorption of a substance is usually manifested as a combination of multiple adjacent bands. A 1×5 convolution can extract the intensity variation trend within this range, which is more comprehensive than single-band analysis.

[0036] In one implementation, normalization (Min-Max normalization) is used to address systematic biases in the spectral acquisition process and ensure data comparability. Then, max pooling is performed on the normalized features to preserve the peak differences of the feature bands, providing a basis for subsequent feature extraction.

[0037] In one embodiment, the improved residual block operates by: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the second convolutional features; The activated features are obtained by batch processing the second convolutional features and then inputting them into the ReLU activation function. The activation features are convolved using a 3×3 convolution operation to obtain the third convolutional feature. The input features are subjected to a 1×3 convolution operation to obtain the fourth convolution feature. The third and fourth convolution features are then added element-wise to obtain the output features of the preprocessing layer.

[0038] In one implementation, a 1×1 convolution is performed on the input features to integrate absorption information from different bands. The characteristic absorption of the same substance may be distributed across multiple discontinuous bands, and the 1×1 convolution can capture the intensity correlation of these bands.

[0039] In one implementation, convolution operations and residual structures are combined to preserve the wavelength continuity of the spectrum, while also mining hidden chemical information through deep networks. Furthermore, multi-dimensional features can be fused by element-wise addition of the third and fourth convolution features, thereby enhancing the spectral characterization capability.

[0040] Based on the same inventive concept, this invention also provides an online near-infrared spectroscopy detection device for tobacco based on a linear filter. See also Figure 2 , Figure 2 A schematic diagram of a tobacco online near-infrared spectroscopy detection device based on a linear filter, provided in an embodiment of the present invention, includes: The original near-infrared band acquisition module is used to emit near-infrared light into the target area and obtain multiple near-infrared bands through a linear filter. The reflected light of each near-infrared band is collected to obtain the original near-infrared band set. The preprocessing module is used to preprocess the original near-infrared band of the target to obtain the target's original near-infrared band; the original near-infrared band of the target is any one of the original near-infrared bands in the set of original near-infrared bands. The feature extraction module is used to extract features from the near-infrared band of the target to obtain target sub-features; The detection result determination module is used to construct a target feature vector based on all target sub-features, and then search a preset database based on the target feature vector to obtain the detection result.

[0041] The present invention provides an online near-infrared spectroscopy detection device for tobacco based on a linear filter. This device acquires multiple near-infrared bands through a linear filter, and obtains the combination and overtone absorption information of the molecules corresponding to different bands. After preprocessing different near-infrared bands, feature extraction is performed, thereby improving the prediction of different chemical components in tobacco and increasing the efficiency of online tobacco detection.

[0042] In one embodiment, the preprocessing module includes: The baseline correction value determination module is used to perform a linear approximation fit on the original near-infrared band to obtain the target straight line equation, and determine the baseline correction value of the original near-infrared band based on the straight line equation; The noise spectrum extraction module is used to extract noise from the near-infrared band based on the baseline correction value to obtain the noise spectrum; The near-infrared correction module is used to correct the original near-infrared band according to the noise spectrum to obtain the target near-infrared band.

[0043] In one embodiment, the feature extraction module includes: The first preprocessing feature determination module is used to substitute the target near-infrared band into the preprocessing layer to obtain the first preprocessing feature; The second residual feature generation module is used to continuously substitute the first preprocessed feature into two residual blocks to obtain the first residual feature, and substitute the first residual feature into the improved residual block to obtain the second residual feature; The fourth residual feature generation module is used to substitute the second residual feature into the residual block to obtain the third residual feature, and to substitute the third residual feature into the improved residual block to obtain the fourth residual feature. The target sub-feature generation module is used to perform convolutional pooling operations on the fourth residual feature to obtain the target sub-feature.

[0044] In one embodiment, the first preprocessing feature determination module includes: The first convolutional feature determination module is used to perform a 1×5 convolution operation on the target's near-infrared band to obtain the first convolutional feature. The first preprocessing feature generation module is used to normalize the first convolutional features and then feed them into the max pooling layer to obtain the first preprocessed features.

[0045] In one embodiment, the improved residual block includes: The feature acquisition module is used to acquire input features and perform a 1×1 convolution operation on the input features to obtain the second convolutional features. The activation module is used to batch process the second convolutional features and then input them into the ReLU activation function to obtain activated features; The convolution module is used to perform a 3×3 convolution operation on the activation features to obtain the third convolution feature; The output module is used to perform a 1×3 convolution operation on the input features to obtain the fourth convolution feature, and to add the third and fourth convolution features element by element to obtain the output features of the preprocessing layer.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for online near-infrared spectroscopy detection of tobacco based on a linear filter, characterized in that, The method includes: Near-infrared light is emitted toward the target area and passed through a linear filter to obtain multiple near-infrared bands. The reflected light from each near-infrared band is collected to obtain the original near-infrared band set. The target's original near-infrared band is obtained by preprocessing the original near-infrared band; the original near-infrared band is any one of the original near-infrared bands in the set of original near-infrared bands. Feature extraction is performed on the near-infrared band of the target to obtain target sub-features; A target feature vector is constructed based on all target sub-features, and the detection result is obtained by searching a preset database based on the target feature vector.

2. The method for online near-infrared spectroscopy detection of tobacco based on a linear filter according to claim 1, characterized in that, Preprocessing the original near-infrared band of the target yields the target's near-infrared band, which includes: The target line equation is obtained by performing a linear approximation fit on the original near-infrared band, and the baseline correction value of the original near-infrared band is determined based on the linear equation. The noise spectrum is obtained by extracting noise from the near-infrared band based on the baseline correction value. The target near-infrared band is obtained by correcting the original near-infrared band based on the noise spectrum.

3. The method for online near-infrared spectroscopy detection of tobacco based on a linear filter according to claim 1, characterized in that, Feature extraction of the target in the near-infrared band yields target sub-features, including: Substituting the target's near-infrared band into the preprocessing layer yields the first preprocessing feature; The first preprocessed feature is successively substituted into two residual blocks to obtain the first residual feature, and the first residual feature is substituted into the improved residual block to obtain the second residual feature; Substituting the second residual feature into the residual block yields the third residual feature, and substituting the third residual feature into the improved residual block yields the fourth residual feature; The target sub-feature is obtained by performing convolutional pooling on the fourth residual feature.

4. The method for online near-infrared spectroscopy detection of tobacco based on a linear filter according to claim 3, characterized in that, Substituting the target's near-infrared band into the preprocessing layer to obtain the first preprocessing feature includes: The first convolutional feature is obtained by performing a 1×5 convolution operation on the near-infrared band of the target. The first convolutional features are normalized and then fed into the max pooling layer to obtain the first preprocessed features.

5. The method for online near-infrared spectroscopy detection of tobacco based on a linear filter according to claim 3, characterized in that, The working principle of the improved residual block includes: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the second convolutional features; After batch processing the second convolutional features, the ReLU activation function is input to obtain the activated features; Perform a 3×3 convolution operation on the activation features to obtain the third convolutional feature; The input features are subjected to a 1×3 convolution operation to obtain the fourth convolution feature. The third convolution feature and the fourth convolution feature are then added element-wise to obtain the output feature of the preprocessing layer.

6. A tobacco online near-infrared spectroscopy detection device based on a linear filter, characterized in that, The device includes: The original near-infrared band acquisition module is used to emit near-infrared light into the target area and obtain multiple near-infrared bands through a linear filter. The reflected light of each near-infrared band is collected to obtain the original near-infrared band set. The preprocessing module is used to preprocess the original near-infrared band of the target to obtain the target near-infrared band; the original near-infrared band of the target is any one of the original near-infrared bands in the set of original near-infrared bands; The feature extraction module is used to extract features from the near-infrared band of the target to obtain target sub-features; The detection result determination module is used to construct a target feature vector based on all target sub-features, and to search a preset database based on the target feature vector to obtain the detection result.

7. The online near-infrared spectroscopy detection device for tobacco based on a linear filter according to claim 6, characterized in that, The preprocessing module includes: The baseline correction value determination module is used to perform a linear approximation fit on the original near-infrared band to obtain the target straight line equation, and determine the baseline correction value of the original near-infrared band based on the straight line equation. The noise spectrum extraction module is used to extract noise from the near-infrared band based on the baseline correction value to obtain a noise spectrum; The near-infrared correction module is used to correct the original near-infrared band according to the noise spectrum to obtain the target near-infrared band.

8. The online near-infrared spectroscopy detection device for tobacco based on a linear filter according to claim 6, characterized in that, The feature extraction module includes: The first preprocessing feature determination module is used to substitute the target near-infrared band into the preprocessing layer to obtain the first preprocessing feature; The second residual feature generation module is used to continuously substitute the first preprocessed feature into two residual blocks to obtain the first residual feature, and substitute the first residual feature into the improved residual block to obtain the second residual feature. The fourth residual feature generation module is used to substitute the second residual feature into the residual block to obtain the third residual feature, and substitute the third residual feature into the improved residual block to obtain the fourth residual feature; The target sub-feature generation module is used to perform convolutional pooling operations on the fourth residual feature to obtain the target sub-feature.

9. The online near-infrared spectroscopy detection device for tobacco based on a linear filter according to claim 8, characterized in that, The first preprocessing feature determination module includes: The first convolutional feature determination module is used to perform a 1×5 convolution operation on the near-infrared band of the target to obtain the first convolutional feature. The first preprocessing feature generation module is used to normalize the first convolutional features and then feed them into the max pooling layer to obtain the first preprocessing features.

10. The online near-infrared spectroscopy detection device for tobacco based on a linear filter according to claim 8, characterized in that, The improved residual block includes: The feature acquisition module is used to acquire input features and perform a 1×1 convolution operation on the input features to obtain a second convolution feature. The activation module is used to batch process the second convolutional features and then input them into the ReLU activation function to obtain activated features; The convolution module is used to perform a 3×3 convolution operation on the activation features to obtain a third convolution feature; The output module is used to perform a 1×3 convolution operation on the input features to obtain a fourth convolution feature, and to add the third convolution feature and the fourth convolution feature element by element to obtain the output feature of the preprocessing layer.

Images