Recombinant beef detection system based on Raman scattering imaging technology

Abstract

The present application discloses a recombinant beef detection system based on Raman scattering imaging technology, which relates to the field of food non-destructive detection technology. In this system, a point laser emits incident laser light to irradiate the surface of the beef to be detected. The processing component collects the original Raman scattering image through a Raman spectrometer imager and a camera component, and performs image preprocessing to obtain an image to be recognized, which is input into a beef detection model. The position encoding layer extracts a feature vector containing Raman spectral information features and relative position features of different spatial offset distances from the image to be recognized, and inputs it into the Transformer encoder. The built-in attention mechanism is used to mine and extract the classification features to be detected containing offset dimension weight information. Finally, the multi-layer perceptron head outputs the detection classification result. This system can accurately and efficiently achieve non-destructive detection of recombinant beef, and can automatically learn the weight contribution information between different data features, greatly improving the classification accuracy of the model and having good robustness.

Description

Technical Field

[0001] This application relates to the technical field of non-destructive testing of food, and in particular to a recombinant beef detection system based on Raman scattering imaging technology. Background Art

[0002] The recombinant beef technology can effectively bond the minced beef produced in the processes such as splitting beef carcasses into whole beef. In recent years, in order to seek illegal interests, illegal vendors use the characteristics that recombinant beef is not easily distinguishable from original cut beef, and replace the labels of recombinant beef with those of original cut beef, seriously damaging the interests of consumers. Therefore, it is necessary to effectively detect recombinant beef to distinguish it from original cut beef.

[0003] The polymerase chain reaction (PCR) based on DNA information and the enzyme-linked immunosorbent assay (ELISA) based on specific antibodies are the most sensitive destructive analysis methods. Chromatography-based detection methods provide chemical separation and have the ability of qualitative and quantitative analysis. Gas chromatography (GC) and high performance liquid chromatography (HPLC) are commonly used in the detection scenarios of beef quality. Electrophoresis methods are based on the principle that proteins will be separated from muscle tissues in an electric field to identify the sarcoplasmic proteins present in beef and related products, so as to identify the beef quality. These traditional physical and chemical methods are widely used in beef quality detection and have the potential for recombinant beef detection, but they have disadvantages such as high cost, low efficiency and dependence on professional operations, and cannot meet the existing requirements for recombinant beef detection. Therefore, they cannot be applied to the recombinant beef detection scenario. Summary of the Invention

[0004] In view of the above problems and technical requirements, the applicant of this application proposes a recombinant beef detection system based on Raman scattering imaging technology. The technical solution of this application is as follows:

[0005] A recombinant beef detection system based on Raman scattering imaging technology, the system includes: a point laser, a Raman spectroscopy imager, a camera assembly and a processing assembly. The point laser emits incident laser to irradiate on the surface of the beef to be detected. The Raman spectroscopy imager receives photon signals from a preset scan line where the incident laser irradiates and disperses them into different wavelengths and projects them onto the camera assembly. The camera assembly acquires the original Raman scattering image of the beef to be detected and transmits it to the processing assembly. The original Raman scattering image includes Raman spectral information within a predetermined wavelength range at each spatial offset distance within a predetermined spatial offset range. The spatial offset distance represents the distance between the Raman spectroscopy imager and the irradiation position of the incident laser on the surface of the beef to be detected;

[0006] The processing assembly performs image preprocessing on the original Raman scattering image to obtain an image to be recognized;

[0007] The processing component inputs the image to be recognized into a beef detection model trained based on a Transformer neural network. The beef detection model includes a position encoding layer, a Transformer encoder, and a multi-layer perceptron head. The position encoding layer extracts a feature vector from the image to be recognized and inputs it into the Transformer encoder. The feature vector contains the Raman spectral information features of the image to be recognized and the relative position features of different spatial offset distances. The Transformer encoder uses the built-in attention mechanism to mine and extract the classification features to be classified that contain offset dimension weight information. The offset dimension weight information indicates the weight information of different spatial offset distances. The multi-layer perceptron head outputs a detection classification result based on the classification features to be classified that contain offset dimension weight information, indicating whether the beef to be detected is original cut beef or reconstituted beef.

[0008] In a further technical solution thereof, in the beef detection model, the method for the position encoding layer to extract a feature vector from the image to be recognized includes: extracting curve embedding features from the Raman spectral information of the image to be recognized in the wavelength dimension, performing position encoding on the curve embedding features to obtain position information features with the same data specification, and adding the curve embedding features and the position information features according to the corresponding positions to obtain a feature vector.

[0009] In a further technical solution thereof, the method for the position encoding layer to extract curve embedding features from the Raman spectral information of the image to be recognized in the wavelength dimension includes:

[0010] Mapping the Raman spectral information of the image to be recognized at each spatial offset distance from a low-dimensional space to a high-dimensional space to extract the curve embedding features of the image to be recognized.

[0011] In a further technical solution thereof, the method for mapping the Raman spectral information of the image to be recognized at each spatial offset distance from a low-dimensional space to a high-dimensional space includes:

[0012] Dividing the image to be recognized into several original sequences according to the spatial offset distance. Each original sequence contains a local image within the corresponding unit spatial offset range in the image to be recognized, and using the method of linear projection to map the Raman spectral information of each original sequence from a low-dimensional space to a high-dimensional space respectively.

[0013] In a further technical solution thereof, the method for the processing component to perform image preprocessing on the original Raman scattering image to obtain the image to be recognized includes:

[0014] Extracting the region of interest in the original Raman scattering image;

[0015] Extracting the Raman spectral information in the typical band from the extracted region of interest to obtain the image to be recognized.

[0016] A further technical solution thereof is that the method for the processing component to extract the region of interest in the original Raman scattering image includes:

[0017] Remove the Raman spectral information within the spatial offset range where the signal-to-noise ratio in the original Raman scattering image is lower than the signal-to-noise ratio threshold, and remove the Raman spectral information within the wavelength range where the effective data within the remaining spatial offset range is lower than the data volume threshold, and intercept the region of interest in the original Raman scattering image.

[0018] A further technical solution thereof is that the method for the processing component to determine the typical band includes:

[0019] Obtain the sample Raman scattering images of a number of beef samples respectively, where the beef samples include raw cut beef samples and reconstituted beef samples;

[0020] Extract the region of interest from the sample Raman scattering images of each beef sample respectively;

[0021] Extract the mean feature of each band based on the regions of interest of all sample Raman scattering images, and select the typical band from the bands by using the successive projections algorithm in combination with the mean feature.

[0022] A further technical solution thereof is that the training method of the beef detection model includes:

[0023] Obtain the sample Raman scattering images of a number of beef samples respectively, where the beef samples include raw cut beef samples and reconstituted beef samples;

[0024] Perform image preprocessing on each sample Raman scattering image respectively to obtain the image to be recognized;

[0025] Input the images to be recognized of each beef sample and their corresponding sample categories into the Transformer neural network for model training to obtain the beef detection model, where the sample category indicates whether the beef sample corresponding to the image to be recognized is a raw cut beef sample or a reconstituted beef sample; the Transformer encoder in the Transformer neural network automatically learns the weight information of different spatial offset distances during the model training process.

[0026] A further technical solution thereof is that the method for obtaining the sample Raman scattering images of a number of beef samples respectively includes:

[0027] Control the incident laser emitted by the laser to move on the surface of the beef sample. When the incident laser irradiates at each irradiation position on the surface of the beef sample, the processing component collects the sample Raman scattering image of the beef sample at the current irradiation position through the camera component and the Raman spectroscopic imager, and combines different irradiation positions to collect multiple sample Raman scattering images of the beef sample at different irradiation positions.

[0028] A further technical solution is that a Droupout layer is introduced into the beef detection model, and the cross-entropy loss function is selected for classification. The Softmax activation function is introduced into the multi-layer perceptron head to generate the detection classification result. The Adam momentum optimization algorithm is introduced during the training process of the beef detection model, and the cosine annealing learning rate is set.

[0029] The beneficial technical effects of this application are as follows:

[0030] This application discloses a recombinant beef detection system based on Raman scattering imaging technology. Raman scattering images not only have the characteristics of high specificity and high sensitivity of traditional Raman spectroscopy but also have the ability to obtain the tissue characteristics of recombinant beef samples. In traditional modeling methods, feature extraction relies on manual experience, and there is a risk of losing effective information. Moreover, the classical sequence data neural network model does not fully consider the problem that the contributions of each data feature are inconsistent in a specific task, which limits the further improvement of the model accuracy. However, the application based on the Transformer neural network in this application can not only fully mine the features in Raman scattering images but also automatically learn the weight contribution information between different data features, greatly improving the classification accuracy of the model and having good robustness, and can accurately and efficiently perform non-destructive detection of recombinant beef and original cut beef respectively.

[0031] Aiming at the serious data redundancy problem in Raman scattering images, this application effectively reduces the data dimension of Raman scattering images by combining the SPA algorithm with the mean features of Raman scattering images while retaining the effective information in the Raman scattering image data to the greatest extent, reducing the complexity of the model and greatly shortening the training time of the deep learning model. Description of the Drawings

[0032] Figure 1 is the system structure diagram of the recombinant beef detection system in an embodiment of this application.

[0033] Figure 2 is the method flow chart for implementing recombinant beef detection in an embodiment of this application.

[0034] Figure 3 is the schematic diagram of data preprocessing and feature vector extraction in an embodiment of this application.

[0035] Figure 4 is the schematic diagram of the beef detection model training in an embodiment of this application. Detailed Embodiments

[0036] The following further describes the detailed embodiments of this application with reference to the drawings.

[0037] This application discloses a recombinant beef detection system based on Raman scattering imaging technology. Please refer to Figure 1The system structure diagram shown, the system includes a point laser 1, a Raman spectroscopy imager 2, a camera assembly 3 and a processing assembly 4. The beef 5 to be detected is placed on the loading platform, and the point laser 1 emits incident laser light that irradiates the surface of the beef 5 to be detected.

[0038] In order to improve the irradiation quality of the incident laser to achieve point light source irradiation, as Figure 1 shown, the incident laser light emitted by the point laser 1 is focused by the focusing assembly 6 and then irradiates the surface of the beef 5 to be detected. The focusing assembly 6 may specifically include an optical fiber collimator and a band-pass filter. In one embodiment, if the point laser 1 is a 785nm point laser, then the band-pass filter is a 785nm laser band-pass filter. When the focusing assembly 6 is arranged obliquely, the incident angle of the incident laser light is about 30°, and on the surface of the beef 5 to be detected, the diameter of the laser spot is about 1.0mm and the power is 350mW.

[0039] The Raman spectroscopy imager 2 receives photon signals from the preset scan line where the incident laser light irradiates and disperses them into different wavelengths and projects them onto the camera assembly 3. In fact, an imaging lens 7 is also installed at the front end of the Raman spectroscopy imager 2, such as a 50mm focusing lens. Figure 1 The dashed line in the figure is the preset scan line. Specifically, the Raman spectroscopy imager 2 receives photon signals from the preset scan line through a 30mm-wide input slit. The Raman spectroscopy imager 2 uses a prism-grating-prism (PGP) device, which disperses the photon signals into different wavelengths and then projects them onto the camera assembly 3. The camera assembly 3 uses a CCD camera, so that the camera assembly 3 can collect the original Raman scattering image of the beef to be detected and transmit it to the processing assembly 4. The processing assembly 4 can be implemented by a computer or a server with data processing functions.

[0040] The original Raman scattering image obtained by the processing assembly 4 includes Raman spectral information within a predetermined wavelength range at each spatial offset distance within a predetermined spatial offset range. The original Raman scattering image is in the form of a two-dimensional image. In the horizontal offset dimension, it represents the spatial offset distance and covers the predetermined spatial offset range, and in the vertical wavelength dimension, it represents the wavelength and covers the predetermined wavelength range. The spatial offset distance represents the distance of a point on the scan line of the Raman spectroscopy imager relative to the irradiation position of the incident laser light on the surface of the beef to be detected, including the predetermined distances on both sides of the irradiation position. The predetermined spatial offset range and the predetermined wavelength range covered by the original Raman scattering image are determined according to the actual situation.

[0041] Please refer to Figure 2 the flowchart shown, the method executed by the processing assembly 4 includes the following steps:

[0042] The processing assembly performs image preprocessing on the original Raman scattering image to obtain an image to be recognized. The image preprocessing mainly includes the following two parts:

[0043] (1) Extract the region of interest in the original Raman scattering image.

[0044] Remove the Raman spectral information within the spatial offset range where the signal-to-noise ratio in the original Raman scattering image is lower than the signal-to-noise ratio threshold, and remove the Raman spectral information within the wavelength range where the effective data within the remaining spatial offset range is lower than the data volume threshold. Then, intercept the region of interest in the original Raman scattering image. In this way, the bands with obviously no effective information and the data within the spatial offset range with large noise in the original Raman scattering image can be removed.

[0045] (2) Extract the Raman spectral information in the typical bands from the extracted region of interest to obtain the image to be recognized. The typical bands are the most representative bands of the Raman spectral information within all wavelength ranges. The typical bands are several pre-determined wavelength ranges, and how to determine the typical bands will be introduced later. According to the several pre-determined typical bands, extract from the region of interest, and the images corresponding to each typical band form an image as the image to be recognized to achieve the purpose of compressing data.

[0046] Please refer to Figure 3 , assuming that the pre-determined wavelength range of the original Raman scattering image covers 1024 bands, and the pre-determined spatial offset range includes a total of 512 pixels covering the pre-determined ranges on both sides of the irradiation position. The region of interest 31 extracted from the original Raman scattering image removes the useless bands at the front and back segments and the useless offset regions with large noise in the original Raman scattering image. Finally, the region of interest 31 covers a total of 324 bands (593 - 1749 cm -1 ) in the wavelength dimension and covers the regions 20 mm on both sides of the irradiation position in the offset dimension. Extract the Raman spectral information in the typical bands from the region of interest 31 to obtain the image to be recognized 32. The coverage range of the image to be recognized 32 in the offset dimension is the same as that of the region of interest 31, and the wavelength dimension range only includes the Raman spectral information in 22 typical bands.

[0047] The processing component inputs the image to be recognized into the beef detection model trained based on the Transformer neural network. The beef detection model successively includes a position encoding layer, a Transformer encoder, and a multi-layer perceptron head. Among them:

[0048] (a) The position encoding layer extracts feature vectors from the image to be recognized and inputs them into the Transformer encoder. The feature vectors contain the Raman spectral information features of the image to be recognized and the relative position features of different spatial offset distances:

[0049] First, extract the curve embedding feature A of the Raman spectral information from the image to be recognized in the wavelength dimension. Specifically: Map the Raman spectral information of the image to be recognized at each spatial offset distance from a low-dimensional space to a high-dimensional space to extract the curve embedding feature A of the image to be recognized. Mapping from a low-dimensional space to a high-dimensional space can increase the separability of the data.

[0050] The actual operation can be, please combine Figure 3 , divide the image 32 to be recognized into several original sequences 33 according to the spatial offset distance. Each original sequence contains the local image within the corresponding unit spatial offset range in the image to be recognized. The range of the spatial offset distance covered by the unit spatial offset range can be customized. For example, in the Figure 3 example shown, divide the image 32 to be recognized into 9 original sequences 33. Each original sequence 33 covers the unit spatial offset range in the offset dimension and all bands of the image 32 to be recognized in the wavelength dimension. Use the method of linear projection to map the Raman spectral information of each original sequence 33 from a low-dimensional space to a high-dimensional space respectively. After all the original sequences 33 are mapped to the high-dimensional space and stitched together according to the offset dimension, the obtained is the curve embedding feature A. At this time, compared with the image 32 to be recognized, the range covered by the curve embedding feature A in the offset dimension remains unchanged, and it is mapped from the original typical band to a higher-dimensional space within the wavelength range.

[0051] Perform position encoding on the curve embedding feature A to obtain the position information feature B with the same data specification. Here, the position encoding can be performed according to the position encoding method of the conventional Transformer neural network. The position information feature B can represent the relative positions between spatial offsets when the spatial offset distance is used as the sequence change dimension. Add the curve embedding feature A and the position information feature B according to the corresponding positions to obtain a feature vector, and input it into the Transformer encoder.

[0052] (b) The Transformer encoder uses the built-in attention mechanism to mine and extract the feature to be classified containing the offset dimension weight information. The offset dimension weight information indicates the weight information of different spatial offset distances.

[0053] (c) The multi-layer perceptron head outputs the detection classification result according to the feature to be classified containing the offset dimension weight information, indicating whether the beef to be detected is original cut beef or reconstituted beef.

[0054] Before using the beef detection model for classification detection, the processing component also includes the training process of the beef detection model. Please refer to the Figure 4 flowchart shown. The training method of the beef detection model includes the following steps:

[0055] Step 410: Obtain the sample Raman scattering images of a number of beef samples respectively. The beef samples include original cut beef samples and reconstituted beef samples.

[0056] The incident laser emitted by the control point laser moves on the surface of the beef sample. When the incident laser irradiates each irradiation position on the surface of the beef sample, the processing component acquires the sample Raman scattering image of the beef sample at the current irradiation position through the camera component and the Raman spectroscopy imager. Combining different irradiation positions, multiple sample Raman scattering images of the beef sample at different irradiation positions are acquired. The method for obtaining the sample Raman scattering image at each irradiation position is similar to the method for obtaining the original Raman scattering image of the beef to be detected described above, and will not be elaborated here.

[0057] Step 420: Perform image preprocessing on each sample Raman scattering image to obtain the image to be recognized. Similar to the above method, it is also necessary to first extract the region of interest from the sample Raman scattering images of each beef sample respectively. Then, the mean features of each band are extracted based on the regions of interest of all sample Raman scattering images, and the successive projections algorithm SPA is used to select typical bands from the bands in combination with the mean features. Using the selected typical bands, the Raman spectral information under the typical bands is extracted from the regions of interest of the obtained sample Raman scattering images to obtain the image to be recognized. On the other hand, the typical bands are also used for image preprocessing of the beef to be detected.

[0058] Step 430: Input the images to be recognized of each beef sample and their corresponding sample categories into the Transformer neural network for model training to obtain a beef detection model. The sample category indicates whether the beef sample corresponding to the image to be recognized is an original cut beef sample or a reconstituted beef sample. The Transformer neural network includes a position encoding layer, a Transformer encoder, and a multi-layer perceptron head. The operations performed by each layer during training are similar to those of the beef detection model, and will not be elaborated here. During the model training process, the attention mechanism is built into the Transformer encoder in the Transformer neural network, which can not only automatically mine the features of the Raman scattering image, but also automatically learn the weight information of different spatial offset distances during the model training process, and pay attention to the data dimensions that are relatively critical for model establishment in the sequence data, thereby further improving the model performance. A Droupout layer is introduced in the beef detection model to alleviate the overfitting phenomenon, and the cross-entropy loss function is selected for classification. The multi-layer perceptron head introduces the Softmax activation function to generate the detection classification result. The Adam momentum optimization algorithm is introduced during the training process of the beef detection model to accelerate the model convergence speed, and the cosine annealing learning rate is set to optimize the training process.

[0059] The above are only the preferred embodiments of the present application, and the present application is not limited to the above embodiments. It can be understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present application shall be considered to be included within the protection scope of the present application.

Claims

1. A recombinant beef detection system based on Raman scattering imaging technology, characterized in that, The system includes: a dot laser, a Raman spectroscopy imager, a camera assembly, and a processing assembly. The dot laser emits incident laser light that irradiates the surface of the beef to be detected. The Raman spectroscopy imager receives photon signals from a preset scan line where the incident laser light irradiates and disperses them into different wavelengths, which are projected onto the camera assembly. The camera assembly acquires the original Raman scattering image of the beef to be detected and transmits it to the processing assembly. The original Raman scattering image is in the form of a two-dimensional image. In the horizontal offset dimension, it represents the spatial offset distance and covers a predetermined spatial offset range. In the vertical wavelength dimension, it represents the wavelength and covers a predetermined wavelength range. The original Raman scattering image includes Raman spectral information within the predetermined wavelength range at each spatial offset distance within the predetermined spatial offset range. The spatial offset distance represents the distance between the Raman spectroscopy imager and the irradiation position of the incident laser on the surface of the beef to be detected; The processing assembly performs image preprocessing on the original Raman scattering image to obtain an image to be recognized; The processing assembly inputs the image to be recognized into a beef detection model trained based on a Transformer neural network. The beef detection model includes a position encoding layer, a Transformer encoder, and a multi-layer perceptron head. The position encoding layer extracts feature vectors from the image to be recognized and inputs them into the Transformer encoder. The feature vectors contain the Raman spectral information features of the image to be recognized and the relative position features of different spatial offset distances. The Transformer encoder uses the built-in attention mechanism to mine and extract the classification features to be classified containing offset dimension weight information. The offset dimension weight information indicates the weight information of different spatial offset distances. The multi-layer perceptron head outputs a detection classification result based on the classification features to be classified containing offset dimension weight information, indicating whether the beef to be detected is original cut beef or reconstituted beef; In the beef detection model, the method by which the position encoding layer extracts feature vectors from the image to be recognized includes: dividing the image to be recognized into several original sequences according to the spatial offset distance. Each original sequence contains the local image within the corresponding unit spatial offset range in the image to be recognized. The Raman spectral information of each original sequence is respectively mapped from a low-dimensional space to a high-dimensional space by means of linear projection to extract curve embedding features. Position encoding is performed on the curve embedding features to obtain position information features with the same data specification. The curve embedding features and the position information features are added together according to the corresponding positions to obtain the feature vectors.

2. The system according to claim 1, wherein The method by which the processing assembly performs image preprocessing on the original Raman scattering image to obtain an image to be recognized includes: Extracting the region of interest in the original Raman scattering image; Extracting the Raman spectral information in the typical band from the extracted region of interest to obtain the image to be recognized.

3. The system according to claim 2, characterized in that, The method by which the processing assembly extracts the region of interest in the original Raman scattering image includes: Remove the Raman spectral information within the spatial offset range where the signal-to-noise ratio in the original Raman scattering image is lower than the signal-to-noise ratio threshold, and remove the Raman spectral information within the wavelength range where the effective data within the remaining spatial offset range is lower than the data volume threshold, and intercept the region of interest in the original Raman scattering image.

4. The system according to claim 3, characterized in that, The method for the processing component to determine the typical band includes: Obtain the sample Raman scattering images of a number of beef samples respectively, where the beef samples include fresh-cut beef samples and restructured beef samples; Extract the regions of interest from the sample Raman scattering images of each beef sample respectively; Extract the mean features of each band based on the regions of interest of all sample Raman scattering images, and select the typical bands from the bands by using the successive projections algorithm combined with the mean features.

5. The system according to claim 1, characterized in that The training method of the beef detection model includes: Obtain the sample Raman scattering images of a number of beef samples respectively, where the beef samples include fresh-cut beef samples and restructured beef samples; Perform image preprocessing on each sample Raman scattering image to obtain the image to be recognized; Input the image to be recognized of each beef sample and its corresponding sample category into the Transformer neural network for model training to obtain the beef detection model, where the sample category indicates whether the beef sample corresponding to the image to be recognized is a fresh-cut beef sample or a restructured beef sample; the Transformer encoder in the Transformer neural network automatically learns the weight information of different spatial offset distances during the model training process.

6. The system according to claim 5, wherein The method for obtaining the sample Raman scattering images of a number of beef samples respectively includes: Control the incident laser emitted by the point laser to move on the surface of the beef sample. When the incident laser irradiates each irradiation position on the surface of the beef sample, the processing component collects the sample Raman scattering image of the beef sample at the current irradiation position through the camera component and the Raman spectroscopic imager, and combines different irradiation positions to collect multiple sample Raman scattering images of the beef sample at different irradiation positions.

7. The system according to claim 5, characterized in that A Droupout layer is introduced in the beef detection model and the cross-entropy loss function is selected for classification. The multi-layer perceptron head introduces the Softmax activation function to generate the detection classification result. The Adam momentum optimization algorithm is introduced during the training process of the beef detection model and the cosine annealing learning rate is set.

Images