Method for detecting wood and white streak mycodermic muscle defect chicken breast by hyperspectral imaging

Abstract

The application discloses a method for detecting chicken breast with wood and white line symbiotic muscle defects by hyperspectral imaging, and comprises the following steps: acquiring wood meat samples and white line meat samples with different grade gradients; acquiring hyperspectral images of the samples; extracting image texture features; preprocessing spectral data; selecting characteristic wavelengths for the spectrum; measuring the physicochemical quality of the samples; taking the image texture features and the spectral data under the selected characteristic wavelengths as inputs, and taking quality index values as outputs, to establish a CNN model; collecting hyperspectral images of chicken breast with wood and white line symbiotic muscle defects, inputting the image texture features and the spectral information of each pixel point in the multispectral image under the characteristic wavelength points into the optimized CNN model, obtaining comprehensive quality index values of each pixel point, respectively assigning colors to each pixel point, and forming a classified visual image; and the model detection and the visual image can accurately analyze and represent the symbiotic state of the white line and the wood tissue.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for meat quality, specifically relating to a method for detecting chicken breast meat with symbiotic woody and white stripe defects using hyperspectral imaging, that is, a hyperspectral imaging detection method for chicken breast meat with symbiotic woody and white stripe defects. Background Technology

[0002] Chicken, delicious, inexpensive, and tender, is an excellent source of many essential nutrients in the human diet. In 2022, my country's total chicken production reached 14.3 million tons, ranking second among global broiler producers, and its share in meat consumption continues to rise. To meet consumer demand for chicken, the poultry industry has adopted continuous breeding and fast-growing farming technologies that offer high feed efficiency, rapid growth, quick slaughter, and high meat yield. However, this has led to serious muscle defects such as white striping and woody / wooden breast, which threaten the industry's development and show a trend of spreading and escalating. White striped meat exhibits clearly visible white streaks parallel to the fibrous tissue on its surface, and is mainly composed of fat; woody / wooden breast is characterized by necrosis of the pectoral muscle tissue, resulting in a hard, pale texture in the protruding areas, and in severe cases, tissue fluid may seep from the surface. White streaks and woody tissues are two types of muscle defect tissues that coexist with each other with a high probability, but there is no good method to detect the degree of coexistence. At present, the detection and evaluation of woody and white streaks meat is still mainly based on human sensory evaluation, which is highly random, subjective, and does not take into account internal quality. The evaluation criteria urgently need to be improved, and the methods are not accurate and effective enough. From a scientific point of view, there is still no good method to explore the coexistence state of woody and white streaks tissues in symbiotic meat.

[0003] Hyperspectral imaging technology organically integrates spectral and two-dimensional imaging technologies, possessing advantages such as image-spectrum integration, multi-band, and high resolution. It has been widely applied in the field of food quality and safety testing. Taking the optical characteristics of woody and white-grained muscle defects in fresh chicken as a starting point, this method combines scientific fields such as food science, veterinary medicine, optoelectronic sensing, non-destructive testing, and artificial intelligence. It integrates systematic experimental data acquisition, meat quality parameter determination, image-spectrum interactive analysis, deep learning model establishment, and visualization representation to build a detection method for the coexistence of woody and white-grained defects. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for detecting muscle defects in chicken breast with symbiotic relationship between wood and white stripes by hyperspectral imaging, in order to overcome the shortcomings of the existing technology, and to solve the current shortcomings of not being able to accurately analyze and characterize the symbiotic relationship between white stripes and wood tissue. At the same time, it also provides methodological support for the breeding of broiler breeds and the study of the pathogenesis characteristics of muscle defects.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging includes the following steps:

[0007] S1: Obtain a certain number (greater than 300) of single wood-fiber chicken breast samples (referred to as wood-fiber meat) of different grade gradients (normal, moderate, severe, extreme) and single white-striped chicken breast samples (referred to as white-striped meat) of different grade gradients (normal, moderate, severe, extreme);

[0008] S2: Build an indoor hyperspectral imaging detection platform with stable light source illumination to obtain hyperspectral images of chicken breast in each sample described in S1;

[0009] S3: Extract image texture features from the hyperspectral image obtained from S2;

[0010] S4: Perform spectral data preprocessing on the spectral data in the hyperspectral image acquired in S2;

[0011] S5: Perform physicochemical quality measurements on the chicken breast in each sample described in S1 to obtain multiple physicochemical quality index values ​​for each sample.

[0012] S6: Combine the values ​​of multiple physicochemical quality indicators in S5 to obtain the comprehensive quality indicator value;

[0013] S7: Select characteristic wavelengths from the spectrum after S4 preprocessing;

[0014] S8: Use the image texture features and spectral data at the selected feature wavelengths of each individual wood-based chicken breast sample as input to the CNN model, and the comprehensive quality index value as output to establish a wood-based chicken breast CNN model and perform model evaluation and optimization; Use the texture features and spectral data at the selected feature wavelengths of each individual white-striped chicken breast sample as input to the CNN model, and the comprehensive quality index value as output to establish a white-striped chicken breast CNN model and perform model evaluation and optimization.

[0015] S9: Hyperspectral images of chicken breast meat with symbiotic woodiness and white vein defects were acquired. Image texture features were extracted from the acquired hyperspectral images. Spectral information of each pixel in the multispectral image at a characteristic wavelength point was selected from the acquired hyperspectral images. The image texture features and the spectral information of each pixel in the multispectral image at the characteristic wavelength point were input into the optimized CNN models for woody and white vein meat, respectively, to obtain the comprehensive quality index value (woodiness or white vein degree score) for each pixel. Each pixel was assigned a color based on its comprehensive quality index value, with different colors representing different comprehensive quality index values. The colored pixels were then stitched together to form a classification visualization image. Using the hyperspectral image of woody white vein meat as the object, the model was substituted for visualization, realizing the separate visualization of woody tissue and white vein tissue, for observing and studying the symbiotic state of woodiness and white veins.

[0016] As a further improvement to the present invention, S2 specifically refers to:

[0017] S21: Preheat the line scanning hyperspectral system for at least 30 minutes, then acquire hyperspectral images of the chicken breast in each sample described in S1 to obtain the original hyperspectral image R. raw ;

[0018] S22: For the original hyperspectral image R raw Perform black-and-white correction to obtain the corrected hyperspectral image R. norm The calculation formula is as shown in (1). Turn on the light source and wait for the light source illuminance to stabilize before acquiring a hyperspectral image of a white reference plate with a reflectance of 100%. white Then, under stable environmental conditions, the light source was turned off, and the camera lens cap was placed to achieve 0% reflectivity, allowing the acquisition of a hyperspectral image R of the black reference plate. dark For the original hyperspectral image R raw The corrected hyperspectral image R is obtained by performing correction using the following formula. norm ,

[0019]

[0020] S23: The surface of chicken breast has a certain curvature, which causes a large difference in the spectral intensity of each pixel in the hyperspectral image. This is relevant to the black-and-white corrected hyperspectral image R of the chicken breast. norm Curvature correction for each sample was performed using formula (2), and the coefficient of variation (CV) of formula (3) was used to evaluate the quality of spectral correction.

[0021]

[0022]

[0023] In the formula A i It is the spectral vector of a single pixel in a hyperspectral image, where P is the number of elements in the spectral vector (i.e., the number of bands), and A is the number of spectral vectors. i(mean) It is the spectral vector after correction of a single pixel, SD is the standard deviation of the spectral matrix, and MN is the mean of the spectral matrix;

[0024] S24: The hyperspectral image after curvature correction is combined with a band math algorithm to determine the region of interest (ROI) to remove spectral information of unwanted pixels, including background, and extract only the pure sample portion. In ENVI software, the band math command is used to subtract the band with the lowest reflectance (422nm band) from the band with the highest reflectance (695nm band) to obtain a band-operated image. A mask for the entire sample is created using 0 as a threshold for the obtained image. Spectral data of the ROI of the chicken breast tissue portion after masking is extracted and averaged to represent each sample. That is, on the band-operated image, pixels with a threshold of 0 or higher are extracted to generate the ROI. The spectral data of all pixels in the ROI are extracted and averaged to extract the average spectrum corresponding to each sample.

[0025] As a further improvement to the present invention, S3 specifically refers to:

[0026] The hyperspectral image acquired by S2 was processed using the Gray-Level Co-occurrence Matrix (GLCM) method to extract image texture information features including mean, contrast, correlation, energy, homogeneity, variance, dissimilarity, and entropy.

[0027] As a further improvement to the present invention, S4 specifically refers to:

[0028] The spectral data in the hyperspectral image acquired by S2 were preprocessed using multivariate scattering correction, standard normal variable transformation, and detrending methods to eliminate the influence of optical scattering in the spectrum.

[0029] As a further improvement to the present invention, in step S5, the physicochemical quality of chicken breast in each sample described in step S1 is measured to obtain multiple physicochemical quality index values ​​corresponding to each sample, specifically as follows:

[0030] S51. Color determination: Take one point in each of the three regions of head, middle and tail on the side surface of chicken breast bone, and use a spectrophotometer to measure the color L*a*b* values ​​of the three points respectively. The average of the three measurements represents the color of the corresponding sample.

[0031] S52. pH measurement: pH measurement is performed using a pH meter equipped with a spear-shaped probe. Before measurement, the pH meter should be calibrated using calibration buffers of pH=4.0 and pH=7.0. The probe of the pH meter is inserted into the muscle tissue about 1 cm below the head, middle and tail of the chicken breastbone side surface for measurement. The probe is rinsed with deionized water after each measurement. The average of three measurements represents the pH value of the corresponding sample.

[0032] S53. Water-holding capacity determination: Among the water-holding capacity indicators, cooking loss (CL) shows significant differences between woody meat, white-striped meat, and normal meat. Drip loss (DL) is the most accurate and objective indicator of water-holding capacity. Therefore, it is proposed to determine cooking loss and drip loss. Drip loss is determined as the percentage of fluid lost from the muscle system under gravity alone, without external force. A piece of chicken breast is cut from the sample for drip loss determination. The formula for the drip loss rate is:

[0033] DL=100%×(W1-W2) / W1 (4);

[0034] In the formula, DL represents the drip loss rate (%); W1 is the initial sample mass (g); W2 is the sample mass (g) after 48 hours of storage;

[0035] The cooking loss was determined as follows: A piece of chicken breast was cut from the sample. The chicken breast was first cooked, and a handheld digital thermometer with a probe was inserted into the thickest part of the chicken breast for real-time measurement. Cooking was stopped when the measured temperature reached 78℃. The cooked chicken breast was then removed and cooled to room temperature. Surface moisture was absorbed with filter paper, and the cooked sample was weighed. The formula for the cooking loss rate is:

[0036] CL = 100% × (W) f -W c ) / W f (5);

[0037] In the formula, CL represents the cooking loss rate (%); W f Indicates the mass (g) of fresh chicken breast; W c This represents the weight (g) of the chicken breast after steaming or boiling.

[0038] S54. Shear force determination: The WBSF of chicken breast was determined using a texture analyzer. After the cooking loss of chicken breast was measured, two 1.9cm wide strips of chicken breast were taken from each sample. The cutting of the strips should be parallel to the direction of the muscle fibers. After preparing the chicken strips for each group, the strips were placed on the texture analyzer in sequence. The blade cut the strips longitudinally from the middle of the strips, perpendicular to the direction of the muscle fibers. The maximum value of the entire cutting process was recorded as the WBSF value of the strip. The average WBSF value of the two strips represents the shear force of the corresponding sample.

[0039] S55: Nutritional quality determination: For samples showing significant differences in moisture, protein, and fat content, a certain mass of chicken breast is cut from the sample, minced, and packaged in a resealable bag. A thoroughly mixed minced meat sample is weighed, and the fat content is determined using Soxhlet extraction with ether and petroleum ether as extractants, according to national standard GB 5009.6-2016. Another thoroughly mixed minced meat sample is weighed, and the protein content is determined using the Kjeldahl method with copper sulfate as a catalyst, according to national standard GB 5009.5-2016. A certain amount of minced meat is placed in a weighing bottle, its mass recorded, and the bottle is placed in an electric thermostatic drying oven for drying. After drying, the bottle is removed, cooled to room temperature in a desiccator, accurately weighed using a high-precision electronic balance, and then placed back in the drying oven for further drying. The moisture content (%) of the sample is calculated when the change in meat mass after continuous weighing does not exceed 0.1%.

[0040] As a further improvement to the present invention, S6 specifically refers to:

[0041] Pearson correlation analysis and principal component analysis were performed on the multiple physicochemical quality index values ​​corresponding to each sample obtained in S5, including S61: constructing an m×n variable matrix X; S62: standardizing each row of matrix X; S63: calculating the covariance matrix C and finding its eigenvalues ​​and corresponding eigenvectors; S64: arranging the eigenvectors into a matrix from top to bottom according to the magnitude of the corresponding eigenvalues, taking the first column to form a matrix, and finding the column vector with the highest contribution rate as the scoring basis for the fusion of physicochemical indexes of woody flesh and white-grained flesh, thus obtaining the comprehensive quality index value after the fusion of multiple physicochemical quality indexes, that is, the woodiness score or the white-graining score.

[0042] As a further improvement to the present invention, S7 specifically refers to:

[0043] The characteristic wavelengths of the spectrum after S4 preprocessing are selected using the Continuous Projection Algorithm (SPA). This involves using vector projection analysis, projecting the wavelength onto other wavelengths, comparing the magnitudes of the projection vectors, and selecting the wavelength with the largest projection vector as the candidate wavelength. The final characteristic wavelength is then selected based on a correction model. The specific method is as follows:

[0044] Let the initial iteration vector be X. k(0) The number of variables to be extracted is N, and the spectral matrix has I columns;

[0045] Choose any i-th column of the spectral matrix and assign the i-th column of the modeling set to X. i , denoted as X k(0) ;

[0046] Let S be the set of positions of the unselected column vectors:

[0047]

[0048] Calculate X respectively i Projection of the remaining column vectors:

[0049]

[0050] Extracting the spectral wavelength of the maximum projection vector:

[0051]

[0052] Let X i =P X If i∈S, n=n+1, and n<N, then this process is calculated cyclically.

[0053] As a further improved technical solution of the present invention, a simplified CNN model is established using the wavelength selected by SPA. The established model includes a woody meat CNN model and a white-striped meat CNN model. Both the woody meat CNN model and the white-striped meat CNN model include an input layer, a convolutional layer C1, a convolutional layer C2, a pooling layer S3, a convolutional layer C4, a convolutional layer C5, a pooling layer S6, a convolutional layer C7, a convolutional layer C8, a pooling layer S9, a fully connected layer F10, a fully connected layer F11, and an output layer, which are used to predict the scores of woody meat and white-striped meat, respectively.

[0054] As a further improved technical solution of the present invention, each sample in S1 includes a left chicken breast and a right chicken breast.

[0055] In step S2, a hyperspectral image of the left or right chicken breast in each sample described in step S1 is obtained.

[0056] In step S5, the physicochemical quality of the right or left chicken breast in each sample described in step S1 is measured to obtain multiple physicochemical quality index values ​​for each sample.

[0057] As a further improvement of the present invention, S2 also includes building a hyperspectral detection platform;

[0058] The hyperspectral detection platform includes:

[0059] A computer used for analyzing and processing hyperspectral images;

[0060] A dark box, the bottom of which is equipped with a platform, and chicken breast is placed on the platform;

[0061] A hyperspectral imager, which is installed on the inner wall of the top of the dark box, facing the chicken breast on the stage;

[0062] The computer is electrically connected to the hyperspectral imager, and the computer can read and process the information identified by the hyperspectral imager.

[0063] The top of the inner wall of the dark box is equipped with an adjustable halogen tungsten lamp light source, which is placed around the hyperspectral imager and can adjust its illumination angle and light intensity.

[0064] The bottom of the darkroom is equipped with a height-adjustable platform.

[0065] The beneficial effects of this invention are as follows:

[0066] 1. This invention utilizes hyperspectral imaging technology combined with deep learning algorithm modeling to identify and visualize the distribution of chicken breast meat with symbiotic defects of wood and white stripes, replacing traditional sensory evaluation. It has the advantages of being fast, non-destructive, and accurate, and solves the defects of human judgment, such as strong subjectivity, inability to find common distribution characteristics, and inability to invert the intertwined symbiosis in the tissue.

[0067] 2. The hyperspectral imaging combined with deep learning modeling detection method provided by this invention is the first to propose a detection method for symbiotic defects of wood and white veins. It can realize the rapid display of the two types of symbiotic tissues, which is beneficial to the exploration of symbiotic mechanism and provides a theoretical basis. At the same time, it provides important reference for the determination of key common control points, reverse guidance of genetic selection and nutritional regulation.

[0068] 3. This invention can find fingerprint spectral information of woody and white-veined chicken meat, and correctly select the characteristic wavelengths for spectral detection of woody and white-veined chicken meat, which is conducive to reducing the overall application cost of the equipment and developing related instruments towards portability and high-throughput detection and sorting. Attached Figure Description

[0069] Figure 1 This is a flowchart of the hyperspectral imaging detection process for chicken breast with woody and white vein symbiosis in this invention.

[0070] Figure 2 This is a diagram showing the sample types involved in this invention.

[0071] Figure 3 This is a structural diagram of the hyperspectral imaging acquisition system in this invention.

[0072] Figure 4 This is a schematic diagram showing the locations of the physicochemical quality detection and hyperspectral detection of chicken breast in this invention.

[0073] Figure 5 This is a schematic diagram of the image texture features extracted in this invention.

[0074] Figure 6 This is a block diagram of the convolutional neural network model used in this invention.

[0075] Figure 7 This is a schematic diagram of the original hyperspectral image and visualization image of chicken meat in this invention. Detailed Implementation

[0076] To make the methods, solutions, and advantages of the present invention clearer, the present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only for illustrating the present invention and are not intended to limit the scope of the present invention.

[0077] A method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging, the entire process is as follows: Figure 1 As shown, the chicken breast meat detection object involved in the embodiments of the present invention is as follows: Figure 2 As shown, this includes normal meat, woody meat (single woody meat), white-veined meat (single white-veined meat), and woody-white-veined meat (woody-white-veined symbiotic meat). The specific methods include:

[0078] Step a: A large number of chicken breast samples from two types of white-feathered chickens with high and low probabilities of muscle defects were obtained from the slaughterhouse for research. The specific experimental conditions included: (1) obtaining Ross 308 and Cobb 500 white-feathered chickens that were 6 weeks old (42d) under the same feeding conditions; (2) ensuring the same fasting time before slaughter, and the slaughtering process included electric stunning, bleeding, high-temperature scalding and plucking, removing the head and feet, removing the internal organs for cleaning, and air cooling for 60-65 minutes; (3) experienced staff manually deboned and skinned the chicken breasts on both sides. On the boneless chicken breast sorting line, long-term trained personnel combined visual inspection, measurement and palpation to select and grade the chicken breasts. As shown in Table 1, the samples were divided into four grade gradients of single woody meat and single white-striped meat according to the existing grading standards: normal, moderate, severe and extreme. Later, different degrees of woody and white-striped symbiotic meat were selected to study the symbiotic state of the two.

[0079] Table 1. Sample Grading Standards:

[0080]

[0081] m samples were selected from each experimental batch, sealed in airtight bags, and transported back to the laboratory in ice-filled containers. There was no significant difference in the degree of muscle defects between the left and right sides. Left (or right) chicken breast was randomly selected primarily for optical data acquisition, while right (or left) breast was primarily used for phenotypic meat quality index determination. Figure 4 As shown;

[0082] Step b: Build and debug the optical characteristic information acquisition system of the hyperspectral imager, such as... Figure 3 As shown. Its core is a line-scan imaging spectrometer, which consists of spectrometer 2 (Imspector V10E, Specim, Oulu, Finland), a 12-bit dynamic range CCD sensor 1 (SensiCam QE SVGA, Cooke Corp., Auburn Hills, MI, USA), C-Mount objective 3 (XNP1.4 / 17-0303, Schneider Optics, Hauppage, NY, USA), etc. The CCD camera has a resolution of 1376×1040 and can capture hyperspectral images covering 520 bands (368-1024nm). Each band has an image resolution of 688×500 pixels, corresponding to a field of view (FOV) of 115mm (width) × 93mm (height). The spectrometer 2 has an internal slit of 30μm (width) × 14.2mm (length) and a spectral resolution of 2.8nm. It removes low signal-to-noise ratio spectral image regions on both sides (368-400nm, 1000-1024nm), retaining only the hyperspectral image within the 400-1000nm spectral range (473 bands). A schematic diagram of the entire system is shown below. Figure 3 As shown. Other accessories include: two 50W MR16 halogen tungsten lamps 4, an electric translation stage 6 (STGA-10, Newmark Systems, Mission Viejo, Cal.), a computer 7, and acquisition and control software. The electric translation stage has a height-adjustable platform, which is remotely controlled. Chicken breast samples 5 are placed on the platform. The computer 7 is used to process chicken quality grade evaluation information. The platform, spectrometer 2, and halogen tungsten lamps 4 are all located inside a dark box. The spectrometer 2 is mounted on the inner top wall of the dark box, facing the chicken sample on the platform. The computer 7 and spectrometer 2 are electrically connected. The computer 7 can read and process information identified by the hyperspectral imager. The dark box material has good light-shielding properties. An adjustable halogen tungsten lamp 4 is placed on the top of the dark box's inner wall, surrounding the spectrometer 2, and its illumination angle and intensity are adjustable. This system primarily utilizes its visible wavelength to enrich external sensory information and its near-infrared wavelength to enrich internal chemical information to obtain superficial tissue information of the skin side's woody and white-veined flesh. The acquired raw hyperspectral images are then processed...raw Perform black and white correction using the following formula, calculated as shown in (1). Turn on the light source and, when the light source illuminance is stable, acquire a hyperspectral image of a white reference plate with a reflectance of 100%. white Then, under stable environmental conditions, the light source was turned off, and the camera lens cap was placed to achieve 0% reflectivity, allowing the acquisition of a hyperspectral image R of the black reference plate. dark For the original hyperspectral image R raw The corrected hyperspectral image R is obtained by performing correction using the following formula. norm :

[0083]

[0084] Among them, R norm For the corrected diffuse reflectance spectral image data, R raw R represents the number of original diffuse reflectance spectral images of the sample. dark For dark image data, R white This is the diffuse reflection image data of the whiteboard.

[0085] Step c: Correct for uneven illumination and sample curvature. The surface of chicken breast is not completely flat, and the height and angle of the light source can cause uneven illumination, resulting in a gradient in light intensity distribution across the sample surface. To obtain more accurate hyperspectral image information, we plan to correct for uneven illumination and surface curvature to rectify morphological effects. This mainly involves first correcting the illumination of each pixel P in the image at a specific wavelength. λ The spectral intensity values ​​are normalized, and then compared with the sum of the spectral intensity values ​​of all pixels at all wavelengths (n), resulting in the final data cube X. λ It is a hyperspectral image that is independent of illuminance, light source direction and curvature. The specific calculation is shown in the following formula (1).

[0086]

[0087] Step d: Remove background and other unnecessary pixel spectral information from the hyperspectral image after illumination and curvature correction. Subtract the low reflectivity 422nm band image from the high reflectivity 695nm band image to obtain the band operation image. On the band operation image, extract the pixels with a threshold of 0 or higher to generate the region of interest (ROI). Extract the spectral data of all pixels in the ROI and average them into a spectrum. Extract the average spectrum corresponding to each chicken breast sample.

[0088] Step e: The hyperspectral image of chicken breast obtained in step d is preprocessed using the multivariate scattering correction (MSC) method. MSC can effectively eliminate spectral differences caused by different scattering levels, thereby enhancing the correlation between the spectrum and the data. This method corrects baseline shifts and offsets in spectral data using ideal spectra. The specific method is as follows:

[0089] The ideal spectrum is obtained by averaging all spectral data.

[0090]

[0091] Perform a univariate linear regression between the spectrum of each sample and the average spectrum, and solve the least squares problem to obtain the baseline shift and offset for each sample:

[0092]

[0093] The spectrum of each sample is corrected, the calculated baseline shift is subtracted, and then divided by the offset to obtain the preprocessed and corrected spectrum:

[0094]

[0095] Step f: The determination of phenotypic meat quality parameters will be carried out using chicken breast samples from the side opposite to the hyperspectral image acquisition sample. The determination of drip loss and nutritional quality requires destructive cutting of the superficial muscle tissue from the chicken breast after the optical characteristics have been acquired. The meat quality parameters will be measured 24 hours post-slaughter when the final quality has stabilized. Measurement locations are as follows: Figure 4 .

[0096] Color determination: Using the Lab color space proposed by the International Commission on Illumination (CIE L*, a*, and b*) as the reference standard, a Minolta CM-2600d spectrophotometer (Konica Minolta Inc., Ramsey, NJ) was used to read the color values ​​on the bone side of the chicken breast (L* represents lightness, a* represents redness, and b* represents yellowness). The average of three measurements taken from the head, middle, and tail ends was used to represent the color of the entire sample.

[0097] pH determination: A Hach pH meter (H280G) equipped with a spear-shaped probe was used for measurement. The pH meter was calibrated with pH 4.0 and pH 7.0 calibration buffers before measurement. The probe was inserted into the muscle tissue about 1 cm below the head and tail for reading. The probe was rinsed with deionized water after each measurement. The average of three measurements was taken to represent the pH value of the entire sample.

[0098] Related research results have found that cooking loss (CL) among water-holding capacity indicators has significant differences among woody meat, white-striped meat and normal meat, while drip loss (DL) is the most accurate and objective indicator to describe water-holding capacity. Therefore, it is proposed to carry out the measurement of cooking loss and drip loss.

[0099] Determination of drip loss: Under the condition of gravity alone without external force, the percentage of fluid loss from the muscle protein system within 48 hours was determined. The initial sample mass was 30g, and the following formula was used for calculation:

[0100]

[0101] In the formula, DL represents the drip loss rate (%); W1 is the initial sample mass (g); and W2 is the sample mass (g) after 48 hours of storage.

[0102] Determination of cooking loss: The sample was steamed in an oven. A handheld digital thermometer with a probe was inserted into the thickest part of the meat to measure the temperature at any time. When the internal temperature reached 78°C, the cooking was complete. The chicken breast was removed and cooled to room temperature. Excess moisture on the surface was absorbed with filter paper. The cooked chicken breast was weighed and the weight was calculated using the following formula.

[0103]

[0104] In the formula, CL represents the cooking loss rate (%); W f Indicates the mass (g) of fresh chicken breast; W c This represents the weight (g) of the chicken breast after steaming or boiling.

[0105] Shear force determination: The WBSF of chicken breast was determined using a texture analyzer. After the cooking loss of chicken breast was measured, two 1.9 cm wide strips of chicken breast were taken from each sample. The cutting of the strips should be parallel to the direction of the muscle fibers. After preparing the chicken strips for each group, the strips were placed on the texture analyzer in sequence. The blade cut the strips longitudinally from the middle of the strips, perpendicular to the direction of the muscle fibers. The maximum value of the entire cutting process was recorded as the WBSF value of the sample. The average of the two shear forces was used to represent the shear force of the entire sample.

[0106] Nutritional quality determination: For chicken breasts with significant differences in content, moisture, protein, and fat were determined. A certain weight of chicken breast was minced and packaged in a resealable bag. A thoroughly mixed sample of the minced meat was weighed, and the fat content was determined using Soxhlet extraction with ether and petroleum ether as extraction solvents, according to GB 5009.6-2016. Another thoroughly mixed sample of minced meat was weighed, and the protein content was determined using the Kjeldahl method with copper sulfate as a catalyst, according to GB 5009.5-2016. A certain amount of minced meat was placed in a weighing bottle, its mass recorded, and the bottle was placed in an electrically heated constant-temperature drying oven for drying. After drying, the bottle was removed, cooled to room temperature in a desiccator, accurately weighed using a high-precision electronic balance, and then placed back in the drying oven for further drying. The moisture content (%) of the sample was calculated when the change in meat mass from continuous weighing did not exceed 0.1%.

[0107] Step h: Conduct Pearson correlation analysis on the above-measured meat quality parameters (color, pH value, drip loss, cooking loss, shear force, etc.). Numerically, this is equal to the product of the covariance (cov(X, Y)) and the standard deviation (σX, σY). Stronger correlation indicates that the parameters tend to be collinear. Further, perform KMO and Bartlett tests on the original meat quality parameters, and calculate the strength of partial correlation or independence between variables based on the test coefficients to determine if factor analysis is suitable. Then, extract principal component features and rotate them using the maximum equilibrium value method, maximum variance method, and maximum fourth order method. Compare the results before and after rotation and select the factor with the largest contribution rate in the analysis as the main comprehensive evaluation information reflecting the degree of muscle defects from a meat quality perspective, as shown in the following formula:

[0108]

[0109] In the formula, X represents a meat quality parameter, a represents factor loadings, F represents common factors, and ε represents specific factors. F and ε are uncorrelated and have a covariance of 0, while F1…F m They are also unrelated, meaning they are sets of mutually perpendicular factors.

[0110] Step i: Extract image texture information features, including mean, contrast, correlation, energy, homogeneity, variance, dissimilarity, and entropy, from the image within the region of interest determined in step d using the Gray-Level Co-occurrence Matrix (GLCM) method. Figure 5 As shown, Figure 5 The image texture information features of normal meat and meat with white streaks or woody texture are shown below:

[0111] Find the mean:

[0112]

[0113] Calculate the contrast:

[0114]

[0115] Calculate the correlation:

[0116]

[0117] Calculate the energy:

[0118]

[0119] Find homogeneity:

[0120]

[0121] Calculate the variance:

[0122]

[0123] Find the dissimilarity:

[0124]

[0125] Calculate entropy:

[0126]

[0127] In the formula, g(i,j) represents the gray-level co-occurrence matrix, X is the number of columns in the gray-level co-occurrence matrix, and Y is the number of rows in the gray-level co-occurrence matrix, where:

[0128]

[0129]

[0130]

[0131]

[0132] Step j: The extracted spectrum obtained after preprocessing in step e is used to select feature wavelengths using the Continuous Projection (SPA) algorithm. Utilizing vector projection analysis, the wavelength is projected onto other wavelengths, the magnitudes of the projection vectors are compared, and the wavelength with the largest projection vector is selected as the candidate wavelength. Then, the final feature wavelength is selected based on the correction model. The specific method is as follows:

[0133] Let the initial iteration vector be X. k(0) The number of variables to be extracted is N, and the spectral matrix has I columns.

[0134] Choose any i-th column of the spectral matrix and assign the i-th column of the modeling set to X. i , denoted as Xk(0) .

[0135] Let S be the set of positions of the unselected column vectors:

[0136]

[0137] Calculate X respectively i Projection of the remaining column vectors:

[0138]

[0139] Extracting the spectral wavelength of the maximum projection vector:

[0140]

[0141] Let X i =P X If i∈S, n=n+1, and n<N, then this process is calculated cyclically.

[0142] Step k: Data analysis and model building: The complex physiological characteristics of wood and white veins lead to nonlinear changes in their optical properties. In order to build an accurate detection model using hyperspectral information, based on the spectral fingerprint data of wood and white veins of all grades, and combined with spectral and image preprocessing methods, we studied targeted chemometric algorithms to establish a qualitative and quantitative model.

[0143] A 13-layer one-dimensional CNN (1D-CNN) was constructed, including an input layer, convolutional layers C1, C2, pooling layers S3, C4, C5, S6, C7, C8, S9, fully connected layers F10 and F11, and an output layer. The structure and process of the 1D-CNN are as follows: Figure 6As shown. Compared to traditional CNNs, a flattened layer is added to the fully connected layer F10 to avoid the problem that multi-dimensional output features cannot be directly connected to the fully connected layer F11. This framework allows for input of fused data including spectral information (average spectrum of the target region) and image information (extracted texture features). Since the network structure cannot recognize labels of type 0, 1, 2, 3 (degree of muscle defect), the category labels need to be converted to one-hot encoding, storing the n-bit state in an n-bit register in the form of 0s and 1s. A small convolutional kernel is used based on the initial size of the input dimension. Unlike traditional CNNs, the input of a one-dimensional CNN is one-dimensional, so the convolutional and pooling layers are also one-dimensional. A one-dimensional convolutional layer (Conv1D) with a size of 3×1 and a stride of 1 is used to extract features. A 3×1 one-dimensional pooling layer is added after convolution, using max pooling to retain as many main features as possible. Valid padding is used to pad the convolution and pooling operations. Considering that the dimension after feature selection is low and subsequent calculations are not possible, the "padding=VALID" for rounding down in convolution is changed to "padding=SAME" for rounding up. The calculation formulas for both are as follows:

[0144]

[0145]

[0146] In the formula, f is the kernel size and s is the stride size.

[0147] Subsequently, the Softmax function is used as the activation function for the fully connected layers, and the sample label is determined by comparing the probability magnitudes. The Softmax function is chosen instead of the Euclidean radial basis function used in LeNet-5 because it is suitable for multi-class classification problems and has better probability distribution characteristics. This study chooses Tanh instead of ReLU as the activation function because ReLU transforms feature values ​​less than zero to zero, which would lose some original information about the samples and increase the sparsity of the network. The classification cross-entropy loss function and the adaptive moment estimation (Adam) optimizer are used to improve training. The parameters alpha, beta_1, beta_2, and epsilon of the Adam optimizer are set to default values ​​of 0.001, 0.9, 0.999, and 1×10⁻⁶, respectively. -8 .

[0148] Step 1: Model Evaluation and Optimization: The correlation coefficient (r) between the actual and predicted values ​​is an important indicator for evaluating the predictive ability of a model. A higher correlation coefficient indicates more accurate predictions, a smaller difference between the predicted and actual values, and a stronger model. The formula for calculating the correlation coefficient is as follows:

[0149]

[0150] In the formula, x and y are two variables representing the actual value and the predicted value, respectively. and The correlation coefficient represents the average of the predicted and actual values, where N is the total number of samples. In spectral analysis, the correlation coefficient is also commonly used to characterize the relationship between wavelength and corresponding index. The absolute value of the correlation coefficient is used to select the characteristic band with the highest correlation. The correlation coefficient ranges from -1 to 1.

[0151] The root mean square error (RMSE) is primarily used to evaluate the model's fit; the closer the correlation coefficient is to 0, the better the model's fit. RMSE can be further divided into the root mean square error of the calibration set (RMSEC), the root mean square error of the cross-validation set (RMSECV), and the root mean square error of the prediction set (RMSEP). The general formula for calculating RMSE is as follows:

[0152]

[0153] In the formula, X i Y represents the measured value of the sample. i is the predicted value for the sample, and n is the number of samples in the model.

[0154] Step m: Visualization of the coexistence of wood and white grain tissue: Hyperspectral images are data cubes that combine spectral and image information. Spectral information can be mapped one-to-one with pixels in the image. Hyperspectral images have a huge advantage in displaying spatial distribution information. The established model can be applied back to hyperspectral / multispectral images and predict the quality index value of each pixel in the sample. After establishing prediction models for single woody succulents (referred to as the woody succulent CNN model) and single white-grained succulents (referred to as the white-grained CNN model), different degrees of woody-white-grained symbiotic succulents from step a are selected as objects. After obtaining hyperspectral images and selecting regions of interest, the image texture features and spectral information of each pixel in the multispectral image at the selected feature wavelength points are obtained as inputs to the single woody succulent prediction model and the single white-grained succulent prediction model. The output value is the pixel category predicted after being substituted into the model, i.e., the comprehensive quality index value (woody or white-grained degree score). Finally, the visualization distribution maps of woody and white-grained succulents are obtained. Finally, the appropriate color scale range and scale are adjusted to make the visualization images clearer, resulting in a colored visualization image of the symbiotic state of woody-white-grained succulents, as shown below. Figure 7 As shown.

[0155] The scope of protection of this invention includes, but is not limited to, the above embodiments. The scope of protection of this invention is defined by the claims. Any substitutions, modifications, or improvements to this technology that are easily conceived by those skilled in the art fall within the scope of protection of this invention.

Claims

1. A method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging, characterized in that, Includes the following steps: S1: Obtain a certain number of single wood-based chicken breast samples of different grade gradients and single white-striped chicken breast samples of different grade gradients; S2: Obtain hyperspectral images of chicken breast in each sample described in S1; S3: Extract image texture features from the hyperspectral image obtained from S2; S4: Perform spectral data preprocessing on the spectral data in the hyperspectral image acquired in S2; S5: Perform physicochemical quality measurements on the chicken breast in each sample mentioned in S1 to obtain multiple physicochemical quality index values ​​for each sample; specifically: S51. Color determination: Take one point from each of the three regions of head, middle and tail on the side surface of chicken breast bone, and use a spectrophotometer to measure the color L*, a* and b* values ​​of the three points respectively. The average of the three measurements represents the color of the corresponding sample. S52. pH measurement: pH measurement is performed using a pH meter equipped with a probe. Before measurement, the pH meter should be calibrated using calibration buffers with pH=4.0 and pH=7.

0. The probe of the pH meter is inserted into the muscle tissue about 1 cm below the head, middle and tail of the chicken breastbone side surface for measurement. The probe is rinsed with deionized water after each measurement. The average of three measurements represents the pH value of the corresponding sample. S53. Conduct water-holding capacity determination: Water-holding capacity determination includes cooking loss determination and drip loss determination; A piece of chicken breast was cut from the sample for drip loss measurement. The formula for the drip loss rate is: ； In the formula, DL represents the drip loss rate; W1 is the initial sample mass; W2 is the sample mass after 48 h. The cooking loss was determined as follows: A piece of chicken breast was cut from the sample. The chicken breast was first cooked, and a handheld digital thermometer with a probe was inserted into the thickest part of the chicken breast for real-time measurement. Cooking was stopped when the measured temperature reached 78℃. The cooked chicken breast was then removed and cooled to room temperature. Surface moisture was absorbed with filter paper, and the cooked sample was weighed. The formula for the cooking loss rate is: ； In the formula, CL represents the cooking loss rate; W f Indicates the quality of fresh chicken breast; W c This indicates the quality of the chicken breast after steaming or boiling; S54. Shear force determination: The Wöhler-Bühler shear force of chicken breast was determined using a texture analyzer. After the cooking loss of chicken breast was measured, two 1.9 cm wide strips of chicken breast were taken from each sample. When taking samples, the cuts of the strips should be parallel to the direction of the muscle fibers. After preparing the chicken strips for each group, the strips were placed on the texture analyzer in sequence. The blade cut the strips longitudinally from the middle of the strips, perpendicular to the direction of the muscle fibers. The maximum value of the entire cutting process was recorded as the Wöhler-Bühler shear force value of the strip. The average of the Wöhler-Bühler shear force values ​​of the two strips represents the shear force of the corresponding sample. S55: Determination of nutritional quality: Cut a certain mass of chicken breast from the sample, mince it, and package it in a resealable bag; weigh a thoroughly mixed minced meat sample, and determine the fat content using ether and petroleum ether as extractants according to the Soxhlet extraction method; weigh another thoroughly mixed minced meat sample, and determine the protein content using copper sulfate as a catalyst according to the Kjeldahl method; take a certain amount of minced meat and put it into a weighing bottle, record the mass, and then put the weighing bottle into an electric thermostatic drying oven for drying. Take it out and place it in a desiccator to cool to room temperature. Weigh it accurately with an electronic balance and then put it back into the drying oven for drying. Calculate the moisture content of the sample when the change in the mass of the meat does not exceed 0.1% after continuous weighing. S6: Combine the multiple physicochemical quality index values ​​from S5 to obtain the comprehensive quality index value; specifically: Pearson correlation analysis and principal component analysis were performed on the multiple physicochemical quality index values ​​corresponding to each sample obtained in S5, including S61: constructing an m×n variable matrix X; S62: standardizing each row of matrix X; S63: calculating the covariance matrix C and finding its eigenvalues ​​and corresponding eigenvectors; 64: Arrange the feature vectors into a matrix from top to bottom according to the magnitude of the corresponding feature values, take the first column to form a matrix, and find the column vector with the highest contribution rate as the scoring basis after the fusion of the physicochemical indicators of woody flesh and white-grained flesh. That is, obtain the comprehensive quality index value after the fusion of multiple physicochemical quality indicators, that is, the woodiness score or the white-grain score. S7: Select characteristic wavelengths from the spectrum after S4 preprocessing; S8: Use the image texture features and spectral data at the selected feature wavelengths of each individual wood-based chicken breast sample as input to the CNN model, and the comprehensive quality index value as output to establish a wood-based chicken breast CNN model and perform model evaluation and optimization; Use the texture features and spectral data at the selected feature wavelengths of each individual white-striped chicken breast sample as input to the CNN model, and the comprehensive quality index value as output to establish a white-striped chicken breast CNN model and perform model evaluation and optimization. S9: Collect hyperspectral images of chicken breast meat with symbiotic woodiness and white vein defects to be tested. Extract image texture features from the collected hyperspectral images. Select spectral information of each pixel in the multispectral image at the characteristic wavelength point from the collected hyperspectral images. Input the image texture features and the spectral information of each pixel in the multispectral image at the characteristic wavelength point into the optimized woody meat CNN model and white vein meat CNN model, respectively, to obtain the comprehensive quality index value of each pixel. Assign color to each pixel according to the comprehensive quality index value of each pixel, and use different colors to represent different comprehensive quality index values. Stitch together the colored pixels to form a classification visualization image.

2. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 1, characterized in that, Specifically, S2 is: S21: Hyperspectral images of chicken breast in each sample described in S1 were acquired using a line-scan imaging spectrometer to obtain the original hyperspectral image R. raw ; S22: For the original hyperspectral image R raw Perform black-and-white correction to obtain the corrected hyperspectral image R. norm ; S23: The hyperspectral image R after black and white correction norm Perform curvature correction; S24: The hyperspectral image after curvature correction is combined with a band mathematical algorithm to determine the region of interest (ROI) to remove the spectral information of unwanted pixels, including the background, and extract only the pure sample portion. The image at the high reflectivity 695 nm band is subtracted from the image at the low reflectivity 422 nm band to obtain the band operation image. On the band operation image, pixels with a threshold of 0 or higher are extracted to generate the ROI. The spectral data of all pixels in the ROI are extracted and averaged to extract the average spectrum corresponding to each sample.

3. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 2, characterized in that, Specifically, S3 refers to: The hyperspectral image acquired by S2 was used to extract image texture information features including mean, contrast, correlation, energy, homogeneity, variance, dissimilarity and entropy using the gray-level co-occurrence matrix method.

4. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 2, characterized in that, Specifically, S4 is: The spectral data in the hyperspectral image acquired by S2 were preprocessed using the multivariate scattering correction method.

5. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 1, characterized in that, Specifically, S7 refers to: The characteristic wavelengths were selected from the spectrum after S4 preprocessing using a continuous projection algorithm.

6. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 1, characterized in that, Both the woody meat CNN model and the white-textured meat CNN model include an input layer, a convolutional layer C1, a convolutional layer C2, a pooling layer S3, a convolutional layer C4, a convolutional layer C5, a pooling layer S6, a convolutional layer C7, a convolutional layer C8, a pooling layer S9, a fully connected layer F10, a fully connected layer F11, and an output layer.

7. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 1, characterized in that, Each sample in S1 includes a left-side chicken breast and a right-side chicken breast. In step S2, a hyperspectral image of the left or right chicken breast in each sample described in step S1 is obtained. In step S5, the physicochemical quality of the right or left chicken breast in each sample described in step S1 is measured to obtain multiple physicochemical quality index values ​​for each sample.

8. The method for detecting muscle defects in chicken breast with symbiotic woodiness and white veins using hyperspectral imaging according to claim 1, characterized in that, S2 also includes building a hyperspectral detection platform; The hyperspectral detection platform includes: A computer used for analyzing and processing hyperspectral images; A dark box, the bottom of which is equipped with a platform, and chicken breast is placed on the platform; A hyperspectral imager, which is installed on the inner wall of the top of the dark box, facing the chicken breast on the stage; The computer is electrically connected to the hyperspectral imager, and the computer can read and process the information identified by the hyperspectral imager. The top of the inner wall of the dark box is equipped with an adjustable halogen tungsten lamp light source, which is placed around the hyperspectral imager and can adjust its illumination angle and light intensity. The bottom of the darkroom is equipped with a height-adjustable platform.

Images