A substance identification system and method based on hyperspectral diffraction fusion imaging

Abstract

This invention provides a material identification system and method based on hyperspectral diffraction fusion imaging, belonging to the fields of hyperspectral imaging and diffraction imaging technology. This invention combines diffraction imaging technology with hyperspectral imaging technology to create a material identification system. This system provides high spectral resolution, enabling detailed spectral analysis of target objects (microorganisms or micron-sized particles), thereby identifying and distinguishing very subtle spectral and morphological differences. The microbial identification system of this invention not only improves resolution and speed but also achieves multi-dimensional characterization at the single-particle scale, simultaneously providing dual information expression of "multi-wavelength variations of the same target" and "differences among multiple targets at the same wavelength." This material identification system significantly improves imaging speed while maintaining high spectral resolution, demonstrating excellent practicality.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral imaging and diffraction imaging technology, specifically relating to a material identification system and method based on hyperspectral diffraction fusion imaging. Background Technology

[0002] Hyperspectral imaging technology reveals the chemical and physical properties of materials by providing continuous spectral information, thereby enabling precise target identification and classification. It has demonstrated unique advantages in fields such as environmental monitoring, precision agriculture, and medical diagnostics. Zhao et al. used hyperspectral technology to achieve non-destructive detection of chlorophyll content in mangrove leaves; He et al. used a hyperspectral imager carried by a UAV to acquire spectral data of surface soil after winter wheat sowing and fertilization in the study area, and determined the nitrogen content of the sampled soil using chemical analysis methods; Chen et al. developed a new low-cost method for detecting leukemia using hyperspectral imaging (HSI) and forward search (FSM). However, traditional hyperspectral imaging systems are limited by spectral resolution and imaging speed, which restricts their application potential in fast-moving scenarios.

[0003] Diffraction techniques excel in improving imaging resolution, especially at the microscale and in high-precision measurements. Matthew P. et al. established a three-dimensional 3D structure in the X-ray diffraction pattern of myofibrils, providing a more accurate method for diagnosing muscle tissue damage. Diffraction techniques are also used to reveal microscopic defects, stress distribution, and grain structure of materials, which is of great significance for the development of new materials. Li et al. used diffraction methods to characterize materials and combined them with molecularly imprinted electrochemical methods to achieve rapid detection of urethane. Debobrata et al. discovered that laser-induced electron diffraction is sensitive to the chirality of the target, opening up a new avenue for solving ultrafast chiral dynamics. With its excellent resolution and broad application prospects, diffraction imaging technology has become an important tool for revealing the microscopic structure of matter, driving cutting-edge research in many scientific fields. However, its ability to acquire spectral information is limited.

[0004] Therefore, there is a need to develop a system that combines diffraction imaging technology with hyperspectral imaging technology. Summary of the Invention

[0005] To address some shortcomings in existing technologies, this invention provides a material identification system and method based on hyperspectral diffraction fusion imaging. This invention combines diffraction imaging technology with hyperspectral imaging technology to create a material identification system. This system provides high spectral resolution, enabling detailed spectral analysis of target objects (microorganisms or micron-sized particles), thereby identifying and distinguishing very subtle spectral and morphological differences. The microbial identification system of this invention not only improves resolution and speed but also achieves multi-dimensional characterization at the single-particle scale, simultaneously providing dual information representation of "multi-wavelength variations of the same target" and "differences among multiple targets at the same wavelength." This material identification system significantly improves imaging speed while maintaining high spectral resolution, demonstrating excellent practicality.

[0006] To achieve the above-mentioned technical objectives, the present invention employs the following technical means:

[0007] This invention first provides a material identification system based on hyperspectral diffraction fusion imaging, wherein the material includes microorganisms or micron-sized particles; the material identification system includes a light source, a hyperspectral diffraction imaging device, and a computer;

[0008] The polychromatic light emitted by the light source is input into the hyperspectral diffraction imaging device through an optical fiber;

[0009] The hyperspectral diffraction imaging device integrates filters, concave reflective gratings, light shields, CMOS sensors, nanoscale displacement stages, and z-axis displacement stages.

[0010] The outer casing of the hyperspectral diffraction imaging device is provided with a filter fixing base for fixing the filter, a grating fixing hole for fixing the concave reflection grating, a z-axis displacement stage connecting hole for connecting the z-axis displacement stage, and a power supply hole for connecting the power supply; the sides and top of the outer casing are provided with light shields.

[0011] The computer is connected and interacts with the CMOS sensor and z-axis displacement stage in the hyperspectral diffraction imaging device via a connecting cable.

[0012] Preferably, the light source includes a halogen lamp light source;

[0013] The optical fiber and the SMA905 flange are bolted together inside the housing of the hyperspectral diffraction imaging device, and the filter is located behind the SMA905 flange.

[0014] Preferably, the concave reflective grating has 200 lines / mm, an effective diameter of 42mm, and a working wavelength range of 200-800nm.

[0015] Preferably, the light shield is provided with a micro-hole adjustment block, a sample layer and a CMOS base in sequence inside, and a CMOS sensor is provided on the CMOS base;

[0016] The light shield is also provided with micro-holes, and the size and position of the micro-holes are controlled by a micro-hole adjustment block.

[0017] Preferably, the lower end of the CMOS base is fixed to the nanoscale displacement stage by bolts.

[0018] Preferably, the nanoscale displacement stage and the z-axis displacement stage are bolted together by a connecting plate.

[0019] Preferably, the connection line between the CMOS sensor and the nanoscale displacement stage and the computer is in the direction of movement of the nanoscale displacement stage.

[0020] The present invention also provides a substance identification method based on hyperspectral diffraction fusion imaging, wherein the substance identification method is implemented based on the above-mentioned substance identification system;

[0021] The substance identification method includes:

[0022] (1) Turn on the light source, connect the CMOS sensor and the nanoscale displacement stage to the computer, then place the sample in the sample layer, adjust the micropore adjustment block to make the diffraction fingerprint of the image acquisition clear, and adjust the height of the z-axis displacement stage to make the micropore position on the spectral focusing line.

[0023] (2) The CMOS sensor acquires diffraction fingerprints in real time through Openmv IDE, and at the same time, the nanoscale displacement stage is controlled by JC Control software to move a certain distance within an interval.

[0024] (3) The data acquired by the CMOS sensor is transmitted to the computer. The computer extracts the spectral data of the image at different wavelengths, integrates them, and then performs characteristic wavelength screening.

[0025] (4) Normalize the data after filtering by characteristic wavelength, input the processed data into the classification model for material identification and classification, and output the classification results.

[0026] Preferably, in step (2), the nanoscale displacement stage is controlled by JC Control software to move a distance of 1 nm wavelength within 5 ms.

[0027] Preferably, in step (3), the computer extracts the spectral data of the image at a wavelength of 400-800 nm;

[0028] The steps for extraction and integration are as follows:

[0029] ;

[0030] The image is viewed as a matrix I with dimensions m × n (rows × columns), where each element I(i, j) represents the light intensity value of that pixel in that wavelength band.

[0031] Preferably, the characteristic wavelength screening in step (3) includes:

[0032] S1. The UVE method (no information variable elimination method) is used to introduce random noise variables into the original spectrum. After establishing a partial least squares model, the stability index of each wavelength is calculated. The maximum absolute value of the noise variable stability index is used as the threshold to remove wavelengths with stability below the threshold.

[0033] S2. The SPA method (continuous projection algorithm) is used to determine the optimal wavelength combination by iteratively projecting vectors and selecting the wavelength with the largest projection vector as the next selected wavelength each time, until a preset number is reached. The goal is to minimize the collinearity between wavelengths.

[0034] S3. CARS (Competitive Adaptive Reweighted Sampling) is used to establish a partial least squares model through Monte Carlo sampling. The absolute value of the regression coefficient is used as the weight. Wavelengths with smaller weights are gradually eliminated using an exponential decay function. Wavelength combinations are optimized through adaptive reweighted sampling. The final feature wavelengths are determined with the goal of minimizing the root mean square error of cross-validation.

[0035] S4. The integrated spectral data is independently screened and extracted using the three methods S1-S3 to obtain three sets of characteristic wavelengths. The total frequency of each wavelength in the three sets is counted, and the 10 wavelengths with the highest frequency are selected as the final combination of characteristic wavelengths. If there are more than 10 wavelengths with the same frequency, cross-validation is used to further optimize the combination.

[0036] Preferably, the selected characteristic wavelengths include: 621 nm, 617 nm, 615 nm, 623 nm, 611 nm, 439 nm, 445 nm, 430 nm, 437 nm, 437 nm, and 609 nm.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention combines hyperspectral imaging and diffraction imaging techniques to create a material identification system based on hyperspectral-diffraction fusion imaging. This system provides high spectral resolution by integrating hyperspectral imaging with diffraction techniques, enabling detailed spectral analysis of target objects and thus identifying and distinguishing very subtle spectral and morphological differences. Compared to existing methods that rely solely on a single hyperspectral or diffraction technique, this invention not only improves resolution and speed but also achieves multi-dimensional characterization at the single-particle scale, simultaneously providing dual information on "multi-wavelength variations of the same target" and "differences among multiple targets at the same wavelength."

[0039] The material identification system described in this invention significantly improves imaging speed while maintaining high spectral resolution, overcoming the bottleneck of traditional hyperspectral imaging systems in rapidly acquiring spectral data and the limitations of diffraction techniques in the spectral data range. Furthermore, this invention is applicable to various disciplines; any research task involving the identification and detection of microparticles or cells can be achieved using this device, such as cell viability identification and classification. The observation field of this system is approximately 24 mm. 2 It is approximately 100 times larger than a standard 400-pixel optical microscope. The image sensor has a maximum resolution of 2592*1944 and a pixel size of 1.4μm×1.4μm. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall structure of a material identification system based on hyperspectral diffraction fusion imaging.

[0041] Figure 2 This is a schematic diagram of the shield structure of a material identification system based on hyperspectral diffraction fusion imaging.

[0042] Figure 3 This is a schematic diagram of the parameters of a concave reflection grating in a material recognition system based on hyperspectral diffraction fusion imaging.

[0043] Figure 4 This is a spectral image of leaf rust spores collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0044] Figure 5 This is a spectral image of stripe rust spores collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0045] Figure 6 This is a spectral image of polystyrene microspheres acquired by a material identification system based on hyperspectral diffraction fusion imaging.

[0046] Figure 7 This is the average spectrum of three types of spores collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0047] Figure 8 It is a fingerprint image of the same sample at different wavelengths collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0048] Figure 9 It is a fingerprint image of different samples at the same wavelength collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0049] Figure 10 This is a result of feature wavelength screening based on data collected by a material identification system based on hyperspectral diffraction fusion imaging; the left image is a bar chart of feature importance levels, and the right image is a global feature importance wavelength distribution map.

[0050] Figure 11 This is the final result image of the classification of data collected by a material identification system based on hyperspectral diffraction fusion imaging.

[0051] Figure label:

[0052] 1- Z-axis displacement stage; 2- connecting plate; 3- nanometer-scale displacement stage; 4- housing; 5- CMOS base; 6- CMOS sensor; 7- light shield; 8- micro-orifice adjustment block; 9- concave reflective grating; 10- filter; 11- SMA905 flange; 12- optical fiber; 13- light source; 14- computer; 15- sample layer Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0054] Example 1: A Material Identification System Based on Hyperspectral Diffraction Fusion Imaging

[0055] like Figure 1-3 As shown, the hyperspectral diffraction fusion imaging material identification system of the present invention includes a light source 13, a hyperspectral diffraction imaging device, and a computer 14; the polychromatic light emitted by the light source 13 is input into the hyperspectral diffraction imaging device through an optical fiber 12; the hyperspectral diffraction imaging device integrates a filter 10, a concave reflection grating 9, a light shield 7, a CMOS sensor 6, a nanoscale displacement stage 3, and a z-axis displacement stage 1; the outer shell 4 of the hyperspectral diffraction imaging device is provided with a filter fixing base for fixing the filter 10, a grating fixing hole for fixing the concave reflection grating 9, and a z-axis displacement stage connecting hole for connecting the z-axis displacement stage 1; the sides and top of the outer shell 4 are provided with light shields; the computer 14 is connected and interacts with the CMOS sensor 6 and the nanoscale displacement stage 3 in the hyperspectral diffraction imaging device through a connecting cable.

[0056] Specifically, the light source 13 is a halogen lamp, and the polychromatic light emitted by the halogen lamp is input into the hyperspectral diffraction imaging device through the optical fiber 12. The optical fiber 12 is bolted to the SMA905 flange 11 inside the housing 4 of the hyperspectral diffraction imaging device, and the filter 10 is located behind the SMA905 flange 11. The polychromatic light is filtered by the filter 10 to remove light with wavelengths outside the 400-800nm ​​range, preventing interference from stray light.

[0057] Specifically, the housing 4 and concave reflective grating 9 of the hyperspectral diffraction imaging device of this invention are custom-designed. To enable the concave reflective grating 9 to better separate polychromatic light, the separated monochromatic light is directed into the light shield 7. The concave reflective grating 9 is a custom-made part with a grating line count of 200 lines / mm, an effective diameter of 42mm, and a working wavelength range of 200-800nm. This design can separate polychromatic light into monochromatic light of different wavelengths and ultimately focus it into a single light spot. Inside the housing 4, a grating sleeve is built according to the outer diameter of the concave reflective grating 9. The concave reflective grating 9 is embedded inside the grating sleeve and then fastened through the grating fixing hole thread to prevent the concave reflective grating from loosening and causing changes in the position of the spectral focusing line.

[0058] The light shield is designed to block light from other orders of diffraction from illuminating the imaging area of ​​the CMOS sensor, ensuring that the sample layer is only exposed to the coherent light region corresponding to the wavelength of the +1st order diffraction light. It also aims to fix the micro-aperture on the plane perpendicular to the +1st order diffraction light, ensuring that the light spot area diffracted through the micro-aperture covers the entire imaging area of ​​the CMOS sensor. Specifically, the light shield 7 contains, in sequence, a micro-aperture adjustment block 8, a sample layer 15, and a CMOS base 5. The CMOS base 5 houses the CMOS sensor 6. The micro-aperture adjustment block 8 is used to adjust the size and position of the micro-aperture on the light shield, allowing monochromatic light to enter the interior of the light shield 7 through the micro-aperture and then penetrate the sample layer 15. The micro-aperture adjustment block 8, the sample layer 15, and the CMOS base 5 are connected by a transition fit, allowing for fine-tuning of the focal length for different samples. The adjustable distance between the micro-aperture adjustment block 8 and the sample layer 15 is 40-75 mm, and the adjustable distance between the sample layer 15 and the CMOS sensor 6 is 0.5-4 mm.

[0059] Specifically, the lower end of the CMOS base 5 is fixed to the nanoscale displacement stage 3 by bolts; the nanoscale displacement stage 3 and the z-axis displacement stage 1 are fixed by bolts via a connecting plate 2. The nanoscale displacement stage 3 has a precision of 10 nm and a minimum stroke of 50 nm. By operating the displacement stage, the micro-orifice is moved within the spectral plane. The nanoscale displacement stage 3 is made of piezoelectric ceramic. Under the inverse piezoelectric principle, resonance causes the piezoelectric ceramic to undergo minute deformation, thereby achieving precise control. The z-axis displacement stage 1 can be used with a light shield to adjust the position of the micro-orifice, ensuring that the micro-orifice is always at the position of the focused beam line after spectral dispersion.

[0060] Specifically, the connection lines between the CMOS sensor 6 and the nanoscale displacement stage 3 and the computer 14 are in the moving direction of the nanoscale displacement stage 3.

[0061] Example 2: A Material Identification Method Based on Hyperspectral Diffraction Fusion Imaging

[0062] This embodiment uses wheat stripe rust spores (SR), wheat leaf rust spores (LR), and polystyrene microspheres (MS) as experimental objects to verify the feasibility and performance of the material identification system based on hyperspectral diffraction fusion imaging described in Example 1 in the identification and classification of microbial spores.

[0063] (1) Wheat stripe rust spores (SR) and wheat leaf rust spores (LR) were collected from wheat leaves infected with stripe rust or leaf rust. Single spore samples were obtained after culturing and purification. Polystyrene microspheres (MS) were standard-sized microspheres (approximately 20 μm in diameter) used as control experiments.

[0064] (2) The substance identification system based on hyperspectral diffraction fusion imaging in the embodiment is used to detect the sample in step (1). The specific steps are as follows:

[0065] Step 1: ① Turn on the light source and connect the CMOS sensor to the computer;

[0066] ② Place the sample to be observed in the sample layer, adjust the micro-orifice adjustment block to make the diffraction fingerprint of the image acquisition clear, and then adjust the height of the z-axis displacement stage so that the position of the micro-orifice is on the spectral focusing line. At this time, the object-image distance is the optimal imaging distance for the sample.

[0067] ③ Tighten the knobs on the z-axis displacement stage and the micro-hole adjustment block, and wait for further data acquisition.

[0068] Step 2: Remove the top cover of the hyperspectral diffraction imaging device housing, place the prepared sample in the corresponding position in the sample layer of the light shield, and then put the cover plate of the hyperspectral diffraction imaging device housing on.

[0069] Step 3: Connect the JC Control software to the computer to control the nanoscale displacement stage, turn off the computer's Bluetooth module, enable the COM signal module, start the motor shaft, and return the nanoscale displacement stage to its zero position (wavelength 800nm ​​position).

[0070] Step 4: Connect the computer to the CMOS sensor, start the program, and convert the mode to grayscale;

[0071] Step 5: Real-time acquisition of diffraction fingerprints using OpenMV IDE, while simultaneously controlling a nanoscale displacement stage to move a certain distance (1nm wavelength distance) within an interval (5ms) using JC Control software.

[0072] Step 6: Extract the collected fingerprint data frame by frame, and then extract a single diffraction fingerprint (single microorganism) through an algorithm. For the currently selected single diffraction fingerprint or region of interest, crop and extract the diffraction fingerprint of each image under the wavelength of 400-800nm. The 400 images extracted under this region of interest are the diffraction fingerprints of the sample under different wavelengths.

[0073] Step 7: Use data processing software to extract the spectral data of each sample (400 images), and integrate them. The specific extraction formula is as follows;

[0074] ;

[0075] The image is viewed as a matrix I with dimensions m × n (rows × columns), where each element I(i, j) represents the light intensity value of that pixel in that wavelength band.

[0076] Step 8: Randomly divide the integrated spectral data into training and test sets at an 8:2 ratio. Then, use the CARS, UVE, and SPA methods to filter characteristic wavelengths from the data in the training set.

[0077] UVE (Uninformation Variable Elimination) introduces random noise variables into the original spectrum, establishes a partial least squares model, calculates the stability index of each wavelength, and uses the maximum absolute value of the noise variable stability index as a threshold to eliminate wavelengths with stability below the threshold.

[0078] The key formula is as follows:

[0079] Stability metrics:

[0080] Filtering threshold: ,reserve The wavelength;

[0081] Wherein: S j b refers to the stability index of the j-th wavelength variable. j In cross-validation or multiple modeling, the mean(b) is the regression coefficient vector corresponding to the j-th wavelength. j The value of std(b) refers to the average value of the regression coefficients for that wavelength. j ) refers to the standard deviation of the regression coefficient for that wavelength; T refers to the screening threshold, S noise Stability index refers to the stability of random noise variables added artificially.

[0082] SPA (Continuous Projection Algorithm) determines the optimal wavelength combination by iteratively projecting vectors, selecting the wavelength with the largest projection vector as the next selected wavelength each time, until a preset number is reached. The goal is to minimize collinearity between wavelengths.

[0083] The key formula is as follows:

[0084] Projection vector:

[0085] Selection rules:

[0086] Where: Pxj refers to the j-th candidate wavelength x j The projection residual x on the currently selected space j The original vector of the j-th wavelength variable, x k-1 The wavelength vector selected in the previous iteration is referred to as 'k'; the new wavelength number selected in this round is referred to as 'k'.

[0087] CARS (Competitive Adaptive Reweighted Sampling) establishes a partial least squares model through Monte Carlo sampling, uses the absolute value of the regression coefficients as weights, employs an exponential decay function to gradually eliminate wavelengths with smaller weights, and optimizes wavelength combinations through adaptive reweighted sampling, aiming to determine the final feature wavelengths by minimizing the root mean square error of cross-validation.

[0088] The key formula is as follows:

[0089] Exponential decay retention rate:

[0090] Wavelength selection probability:

[0091] Where: r i The proportion of wavelengths retained during the i-th sampling is denoted by 'a', where 'a' is the initial retention proportion and 'k' is the attenuation rate constant. Pj is the probability or weight of the j-th wavelength being selected.

[0092] Three sets of characteristic wavelengths were obtained independently using CARS, UVE, and SPA methods. The total frequency of each wavelength appearing in the three sets was counted, and the 10 wavelengths with the highest frequency were selected as the final combination of characteristic wavelengths. If more than 10 wavelengths had the same frequency, cross-validation was used to further optimize the combination.

[0093] In this step, the characteristic wavelengths obtained by screening are: 621 nm, 617 nm, 615 nm, 623 nm, 611 nm, 439 nm, 445 nm, 430 nm, 437 nm, 437 nm, and 609 nm.

[0094] The above screening steps can efficiently and automatically extract the most discriminative key wavelength combinations from full-spectrum data from different perspectives such as "competitive adaptive reweighting", "minimizing information redundancy" and "global cluster optimization", thereby significantly reducing data dimensionality and improving detection speed and system real-time performance while ensuring recognition accuracy.

[0095] Step 9: Input the feature wavelengths selected from the training set into the SVM model, and adjust the parameters: penalty coefficient C=1, batch size 100, learning rate 1x10⁻¹⁰. -4 By analyzing the features of sample substances and classifying them using SVM, an SVM substance identification classification model was obtained. After training, the accuracy rate of SVM as the classification model exceeded 98%. SVM was chosen as the basic classification model in this step because its algorithm is classic, has strong generalization ability, and performs stably in scenarios with small sample sizes and high-dimensional data. Furthermore, this model has good scalability; if the accuracy is insufficient when applied to other microbial identification scenarios, it can be easily replaced with more advanced classification algorithms (such as ensemble learning, deep learning, etc.) to further improve recognition performance.

[0096] Step 10: Perform validation using the validation set. The overall results are as follows. Figure 11 As shown in the figure, this method can efficiently and automatically extract the most discriminative key wavelength combinations from full-spectrum data from different perspectives, such as "competitive adaptive reweighting," "minimizing information redundancy," and "global cluster optimization." This significantly reduces data dimensionality and improves detection speed and system real-time performance while ensuring recognition accuracy.

[0097] To evaluate the stability and differential expression of this device in spore information acquisition, multiple monomer spectra were acquired for three types of samples (SR, LR, and MS). The results are as follows: Figure 4 , 5 As shown in Figure 6, the three types of samples exhibit consistent morphology and controllable noise levels across the entire spectral range, demonstrating the device's good repeatability and data consistency. Based on this, the average spectra of each type of sample were calculated and compared (see Figure 6). Figure 7 As can be seen from the figure, the average curves exhibit distinguishable overall trend differences across multiple wave segments, providing a usable information basis for subsequent feature extraction and classification.

[0098] In addition, to visually demonstrate the device's information representation capabilities at the diffraction imaging level, the diffraction fingerprints of the same spore at different wavelengths were first displayed (see...). Figure 8 As can be seen from the figure, the diffraction pattern becomes clearer with increasing wavelength, and the diffraction ring radius is positively correlated with the wavelength. Furthermore, comparisons were made between different spores at the same wavelength (see...). Figure 9Within a radius of 25 pixels, leaf rust exhibits the highest light intensity, while microspheres show the lowest. Beyond a radius of 25 pixels, leaf rust spores show the lowest light intensity, while stripe rust spores show the highest, reflecting the differences between the categories. These two sets of results together demonstrate that the combination of spectral and diffraction information can provide an effective characterization for differentiation in two dimensions: "multiple wavelengths of the same target" and "multiple targets of the same wavelength."

[0099] This embodiment further employs a characteristic wavelength screening strategy based on the combined spectral-diffraction data to obtain several representative bands for classification (see...). Figure 10 Using features at selected wavelengths as input, a classification model was constructed, and its performance was evaluated on an independent test set. Results show that the model can effectively distinguish between SR, LR, and MS. Comparisons were made with different feature wavelength selection methods and the model (see Table 1), with the highest accuracy reaching 98.61%. The confusion matrix is ​​shown in [Table 1]. Figure 11 Only one SR sample was misclassified as LR, which shows that the method described in this invention can accurately classify and distinguish microorganisms or micron-sized particles.

[0100] Table 1. Comparison of accuracy rates of different models

[0101] Indicator Model accuracy Accuracy Recall rate F1 score uve-rbf 0.8194 0.8255 0.8194 0.8177 uve-linear 0.8750 0.8855 0.8750 0.8761 uve-poly 0.9583 0.9591 0.9583 0.9584 uve-sigmoid 0.5000 0.2500 0.5000 0.3333 spa-rbf 0.8333 0.8553 0.8333 0.8304 spa-linear 0.9861 0.9868 0.9861 0.9862 spa- poly 0.9861 0.9868 0.9861 0.9862 spa- sigmoid 0.5000 0.2500 0.5000 0.3333 cars-rbf 0.7222 0.7222 0.7222 0.6877 cars- linear 0.9167 0.9178 0.9167 0.9151 cars- poly 0.9028 0.9001 0.9028 0.9007 cars- sigmoid 0.5000 0.2500 0.5000 0.3333

[0102] As can be seen, the material identification system and method based on hyperspectral diffraction fusion imaging described in this invention can stably acquire multiple spectra of three types of samples and form average features of the categories; at the diffraction level, it can simultaneously provide the fingerprint evolution of "the same target changing with wavelength" and the fingerprint differences of "different targets at the same wavelength"; after feature wavelength screening, the classification model achieves an overall accuracy of 98.61% on the test set, verifying the feasibility and effectiveness of this device in classifying and distinguishing microorganisms or micron-sized particles.

[0103] In summary, this invention combines diffraction imaging technology with hyperspectral imaging technology to create a material identification system. This system provides extremely high spectral resolution, enabling detailed spectral analysis of target objects (microorganisms or micron-sized particles), thereby identifying and distinguishing very subtle spectral and morphological differences. The microbial identification system of this invention not only improves resolution and speed but also achieves multi-dimensional characterization at the single-particle scale, simultaneously providing dual information on "multi-wavelength variations of the same target" and "differences among multiple targets at the same wavelength." Furthermore, this material identification system significantly improves imaging speed while maintaining high spectral resolution, demonstrating excellent practicality.

[0104] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A substance identification system based on hyperspectral diffraction fusion imaging, characterized in that, The substance includes microorganisms or micron-sized particles; the substance identification system includes a light source (13), a hyperspectral diffraction imaging device, and a computer (14). The polychromatic light emitted by the light source (13) is input into the hyperspectral diffraction imaging device through the optical fiber (12); The hyperspectral diffraction imaging device integrates a filter (10), a concave reflective grating (9), a light shield (7), a CMOS sensor (6), a nanoscale displacement stage (3), and a z-axis displacement stage (1). The outer shell (4) of the hyperspectral diffraction imaging device is provided with a filter fixing base for fixing the filter (10), a grating fixing hole for fixing the concave reflection grating (9), a z-axis displacement stage connecting hole for connecting the z-axis displacement stage (1), and a power supply hole for connecting the power supply; the side and top of the outer shell (4) are provided with light shields. The computer (14) is connected and interacts with the CMOS sensor (6) and the nanoscale displacement stage (3) in the hyperspectral diffraction imaging device via a connecting cable.

2. The substance identification system according to claim 1, characterized in that, The light source (13) includes a halogen lamp light source; The optical fiber (12) and the SMA905 flange (11) are bolted together inside the housing (4) of the hyperspectral diffraction imaging device, and the filter (10) is located behind the SMA905 flange (11).

3. The substance identification system according to claim 1, characterized in that, The concave reflective grating (9) has 200 lines / mm, an effective diameter of 42mm, and a working wavelength range of 200-800nm.

4. The substance identification system according to claim 1, characterized in that, The light shield is provided with a micro-hole adjustment block (8), a sample layer (15) and a CMOS base (5) in sequence inside. The CMOS base (5) is provided with a CMOS sensor (6). The light shield is also provided with micro-holes, and the size of the micro-holes is controlled by the micro-hole adjustment block (8).

5. The substance identification system according to claim 1, characterized in that, The lower end of the CMOS base (5) is fixed to the nanoscale displacement stage (3) by bolts.

6. The substance identification system according to claim 1, characterized in that, The nanoscale displacement stage (3) and the z-axis displacement stage (1) are fixed by bolts through a connecting plate (2); The connection lines between the CMOS sensor (6) and the nanoscale displacement stage (3) and the computer (14) are in the direction of movement of the nanoscale displacement stage (3).

7. A material identification method based on hyperspectral diffraction fusion imaging, characterized in that, The substance identification method is implemented based on the substance identification system according to any one of claims 1-6; The substance identification method includes: (1) Turn on the light source, connect the CMOS sensor and the nanoscale displacement stage to the computer, then place the sample in the sample layer, adjust the micropore adjustment block to make the diffraction fingerprint of the image acquisition clear, and adjust the height of the z-axis displacement stage to make the micropore position on the spectral focusing line. (2) The CMOS sensor acquires diffraction fingerprints in real time through Openmv IDE, and at the same time, the nanoscale displacement stage is controlled by JC Control software to move a certain distance within an interval. (3) The data acquired by the CMOS sensor is transmitted to the computer, and the computer performs image preprocessing, light intensity feature extraction and feature wavelength screening. (4) Normalize the data after filtering by characteristic wavelength, input the processed data into the classification model for material identification and classification, and output the classification results.

8. The substance identification method according to claim 7, characterized in that, In step (2), the nanoscale displacement stage is controlled by JC Control software to move a distance of 1 nm wavelength within 5 ms.

9. The substance identification method according to claim 7, characterized in that, In step (3), the computer extracts spectral data of the image at wavelengths of 400-800 nm; The steps for extraction and integration are as follows: ； The image is viewed as a matrix I with dimensions m × n (rows × columns), where each element I(i, j) represents the light intensity value of that pixel in that wavelength band.

10. The substance identification method according to claim 7, characterized in that, The characteristic wavelength screening in step (3) includes: S1. The UVE method is used to introduce random noise variables into the original spectrum. After establishing a partial least squares model, the stability index of each wavelength is calculated. The maximum absolute value of the stability index of the noise variable is used as the threshold to remove wavelengths with stability below the threshold. S2. The SPA method is used to iterate through vector projection, selecting the wavelength with the largest projection vector as the next selected wavelength each time, until a preset number is reached, in order to determine the optimal wavelength combination with the goal of minimizing collinearity between wavelengths; S3. A partial least squares model is established using CARS through Monte Carlo sampling. The absolute value of the regression coefficients is used as the weights. Wavelengths with smaller weights are gradually eliminated using an exponential decay function. Wavelength combinations are optimized through adaptive reweighting sampling. The final characteristic wavelengths are determined with the goal of minimizing the root mean square error of cross-validation. S4. The integrated spectral data is independently screened and extracted using the three methods S1-S3 to obtain three sets of characteristic wavelengths. The total frequency of each wavelength in the three sets is counted, and the 10 wavelengths with the highest frequency are selected as the final combination of characteristic wavelengths. If there are more than 10 wavelengths with the same frequency, cross-validation is used to further optimize the combination.

Images