Plastic sorting system based on short wave infrared multispectral camera imaging

Abstract

The invention provides a plastic sorting system based on short-wave infrared multispectral camera imaging. The plastic sorting system comprises a spectral database establishing module, a key wave band screening module, a multispectral classification model building module and an industrial grade sorting device building module. The spectrum database establishment module is used for providing data support for subsequent wave band screening and model training; the key wave band screening module is used for executing a data preprocessing and fusion type feature extraction method and screening out a key wave band with the highest discrimination degree from original wave bands; the multi-spectral classification model construction module constructs a multi-spectral classification model based on the key wave band with the highest distinction degree, and precise identification of plastic types is achieved; the industrial grade sorting device building module is used for building a short wave infrared multispectral imaging plastic sorting device simulating an industrial production line, and real-time dynamic sorting is achieved. According to the invention, dimensionality reduction of the spectral band is realized reasonably and effectively, so that an expensive hyperspectral camera is replaced, and the hardware cost and complexity of the system are reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial sorting and spectral imaging technology, specifically to a plastic sorting system based on short-wave infrared multispectral camera imaging. Background Technology

[0002] With the widespread use of plastics in daily life and industrial production, the environmental pollution and resource waste caused by waste plastics are becoming increasingly prominent. The key to improving the recycling rate of plastics is to achieve efficient and accurate sorting of plastics. However, traditional plastic sorting methods are mainly manual, which has drawbacks such as low efficiency, high labor costs, and harm to workers' health.

[0003] To address the drawbacks of manual sorting, various automated sorting methods have emerged in existing technologies, including mechanical sorting, electromagnetic sorting, and optical sorting. Among these, optical sorting has become the mainstream development direction due to its advantages such as high automation, fast sorting speed, and environmental friendliness. While Raman spectroscopy, laser-induced fluorescence (LAF), and laser-induced breakdown spectroscopy (DIBS) within optical sorting technologies each have their applications, they all have significant limitations: Raman spectroscopy is susceptible to interference from sample fluorescence, leading to decreased accuracy in thick samples; LAF can only detect substances with fluorescent properties, is sensitive to ambient light, and requires a controlled environment; and DIBS has high equipment and maintenance costs and strict requirements for sample surface flatness.

[0004] Near-infrared hyperspectral imaging (NIHH) technology, as an emerging optical sorting technology, can acquire three-dimensional data cubes of objects (including image and spectral information). It features non-destructive detection and high scanning efficiency, showing potential in the field of plastic sorting. However, existing plastic sorting solutions related to NHIH HSI still have several shortcomings: First, hyperspectral cameras are expensive and the systems are complex, making large-scale application in industrial scenarios difficult. Second, the extraction of spectral features and algorithm development for specific plastic types lack universality, making it difficult to adapt to the diverse materials, complex colors, and varied shapes of plastics in industrial production lines. Third, traditional sorting systems are mostly static detection systems, lacking real-time continuous imaging and dynamic sorting capabilities, and their sorting efficiency cannot meet the batch processing needs of industrial production lines. Fourth, hyperspectral data has high dimensionality and redundancy; direct dimensionality reduction easily leads to feature loss, resulting in insufficient classification robustness. Furthermore, dimensionality reduction methods that rely solely on data-driven approaches lack material chemistry knowledge, making it difficult to ensure the effective extraction of key features.

[0005] Furthermore, in existing industrial sorting systems, traditional visual algorithms are easily affected by appearance factors such as the color and shape of plastics, resulting in low recognition accuracy. Some multispectral sorting schemes suffer from problems such as blind band selection, limited feature extraction, and poor generalization ability of classification models. Moreover, they lack dynamic sorting devices adapted to industrial production lines, making real-time, continuous sorting operations impossible. Therefore, developing a low-cost, high-accuracy plastic sorting system adapted to industrial production lines has significant practical importance and application value. Summary of the Invention

[0006] Purpose of the Invention: To address the problems of high cost, poor versatility, low sorting efficiency, and insufficient classification accuracy in existing plastic sorting technologies, this invention provides a plastic sorting system based on short-wave infrared multispectral camera imaging, aiming to achieve the following objectives:

[0007] (1) By reducing the dimensionality of the spectral bands, a low-cost and high-performance short-wave infrared multispectral imaging system is established to solve the problem that hyperspectral cameras are difficult to apply on a large scale in industry.

[0008] (2) A key band selection method integrating data-driven and materials chemistry knowledge is proposed to achieve accurate differentiation of plastic types and avoid feature loss and decreased classification robustness;

[0009] (3) Construct a multi-dimensional feature extraction and efficient classification model to improve the classification accuracy of plastics of different materials, colors and shapes;

[0010] A dynamic imaging sorting device adapted to industrial production lines was built to achieve real-time, continuous sorting of plastics and meet the needs of industrial batch processing.

[0011] This invention provides a low-cost short-wave infrared multispectral imaging plastic sorting system, including a spectral database establishment module, a key band screening module, a multispectral classification model construction module, and an industrial-grade sorting device construction module;

[0012] The spectral database establishment module is used to execute a method for rapidly acquiring short-wave infrared multi-band spectral data of different types of plastics, providing data support for subsequent band selection and model training;

[0013] The key band selection module is used to perform a data preprocessing and fusion feature extraction method to select four key bands with the highest discriminative power from 224 original bands.

[0014] The multispectral classification model building module constructs a multispectral classification model based on four key bands with the highest discrimination, enabling accurate identification of plastic types.

[0015] The industrial-grade sorting device construction module is used to build a short-wave infrared multispectral imaging plastic sorting device that simulates an industrial production line, enabling real-time dynamic sorting.

[0016] The method for rapidly acquiring short-wave infrared multi-band spectral data of different types of plastics includes the following steps:

[0017] Step a1: Install the shortwave infrared hyperspectral camera as the core acquisition device, install halogen lamps on both sides of the shortwave infrared hyperspectral camera, the halogen lamps provide illumination light in the 900~1700nm band, and then place the white calibration plate that provides the diffuse reflection reference and the black calibration plate that provides the noise reference on both sides of the electric push-broom stage.

[0018] Step a2: Place several plastic samples of the same type on an electric push-broom stage for imaging each time, and obtain spectral data of 224 bands in the range of 900~1700nm for each batch of samples, which is the original brightness value DN (Digital Number) obtained by the internal sensor of the short-wave infrared hyperspectral camera.

[0019] Step a3: Repeat step a2 to obtain spectral datasets for different types of plastics.

[0020] The key band screening module is used to perform the following steps:

[0021] Step b1: Preprocess the spectral dataset to obtain the spectral reflectance dataset;

[0022] Step b2: Perform dimensionality reduction on the spectral reflectance dataset to generate a pseudo-multispectral dataset;

[0023] Step b3 involves performing a second dimensionality reduction on the pseudo-multispectral dataset to ultimately select the four key bands with the highest discriminative power.

[0024] Step b1 includes: First, using LabelMe software to mark the outline of the plastic sample and retain the valid pixels inside, thus obtaining the net dataset for each type of plastic sample; then, using the Interquartile Range (IQR) method to remove outliers, i.e., removing DN values ​​in the range of 0.05 and 0.05. The above and The following data, of which , These are the first, third, and fourth quantiles of the dataset, respectively. ;

[0025] Secondly, combining the spectral data from the white and black calibration plate areas, a normalization operation is performed on all the obtained valid data points, using the following formula: Where D is the new data after normalization. , , The data are the valid data points, the diffuse reflectance reference data within the white calibration plate area, and the noise reference data within the black calibration plate area, respectively. The spectral reflectance dataset is obtained based on D.

[0026] Step b2 includes: based on the band and transmittance response curves of the candidate bandpass filter products, performing integral calculations on the data segmented according to different band intervals, transforming the original data into 42-dimensional pseudo-multispectral data, as shown in the formula:

[0027] ,

[0028] in, For wavelength The corresponding filter transmittance coefficient, For wavelength The nth dimension data (reflectivity value) corresponding to the band in question. wavelength infinitesimal increment, These are the lower and upper wavelength boundaries (i.e., the t-th dimension) corresponding to the t-th new segmented band interval, respectively. Let t be the spectral reflectance value of the new spectral dimension after one dimensionality reduction, according to We obtain a dimensionality-reduced spectral reflectance dataset, where Z represents the set of integers.

[0029] Step b3 includes:

[0030] Step b3-1: Based on a dimensionality-reduced spectral reflectance dataset (42 dimensions), randomly select a subset of bands as the initial feature set. Use K-means clustering to classify the samples on this initial feature set, and calculate the ANOVA-F value (reflecting the inter-group differences of the feature subset under different types of plastics) for each attempt to verify the significance of the selected bands.

[0031] Step b3-2: Repeat step b3-1 and use a brute-force recursive method to solve for the optimal band combination, and select 6 candidate bands.

[0032] Step b3-3 involves analyzing the molecular structure and characteristic functional groups of various types of plastics across all candidate wavelengths, based on their chemical properties. Simultaneously, a comparative verification of differences is conducted, ultimately selecting four key wavelengths with the highest discriminative power: 1140nm, 1200nm, 1245nm, and 1310nm, for subsequent multispectral image classification.

[0033] The multispectral classification model construction module performs the following steps:

[0034] Step c1: Select a short-wave infrared multispectral camera with an InGaAs detector as the core component (the cost is much lower than the hyperspectral camera mentioned above). Fix the bandpass filters of the four key bands with the highest discrimination in front of the short-wave infrared multispectral camera so that the camera, filters and plastic sample are on the same horizontal plane. At the same time, place the white calibration plate and the black calibration plate on the left and right sides of the sample.

[0035] Step c2: Use a camera to take a horizontal picture of the plastic sample, while adjusting the height of the halogen lamp and the angle of the lampshade baffle so that the angle between the incident plane of the light and the horizontal plane is... Values ​​in Between these steps, the camera exposure time was adjusted according to the distance, and different types of plastic samples were photographed sequentially to obtain one set of data. Then, the bandpass filters with different center wavelengths were changed, and the photographing was repeated to obtain four sets of data in the end.

[0036] Step c3: Perform the preprocessing of step b1 on the multispectral (note: different from hyperspectral, this is 1D data) dataset of all types of plastic samples obtained in step c2 under 4 different bandpass filters to obtain the normalized spectral reflectance dataset.

[0037] Step c4: Using the normalized spectral reflectance dataset as the base feature, calculate the following spectral enhancement features:

[0038] Original spectral eigenvectors ;in The spectral reflectance value under characteristic band i;

[0039] inter-spectral ratio ;in , These are the spectral reflectance values ​​under characteristic band i and characteristic band j, respectively. The ratio between characteristic bands i and j was used to obtain a total of 6 sets of characteristic data.

[0040] Spectral similarity characteristics ; Based on the similarity features of characteristic bands i and j, a total of 6 sets of feature data were obtained;

[0041] Finally, a 17-dimensional spectral enhancement feature vector dataset was obtained. ;in They are respectively The collection of all data The collection of all data Assign a number to the type of plastic it belongs to (different types of plastics are mapped to different numbers);

[0042] Step c5: Based on the 17-dimensional spectral enhanced feature vector dataset, construct a chemical property-driven hierarchical decision tree (HDT) classification model to replace the traditional single-level decision tree (DT).

[0043] In step c5, the chemical property-driven hierarchical decision tree (HDT) classification model specifically includes:

[0044] First level: Separate polypropylene (PP) and polyethylene (PE) from other polymers to obtain a subset of polypropylene (PP) and a subset of polyethylene (PE). Utilize the chemical properties of both polypropylene (PP) and polyethylene (PE) as non-polar polyhydrocarbons, and the characteristic absorptions generated by their carbon-hydrogen bond (C–H) skeleton vibrations in the near-infrared spectrum, which have obvious spectral distinctions with polymers containing polar groups or isomeric skeletons, to achieve classification.

[0045] The second level: In non-polypropylene (PP) and polyethylene (PE) samples, polyvinyl chloride (PVC) is identified separately to obtain a subset of PVC. The characteristic absorption differences caused by the chlorine-containing carbon-chlorine bond (C–Cl) chemical structure of PVC are used to achieve accurate differentiation.

[0046] The third level: The remaining samples are divided into aromatic and non-aromatic polymers to obtain aromatic subsets and non-aromatic subsets. Based on the combined vibrations and overtones generated by the conjugation or polar groups of aromatic rings, ester groups, and amide groups, the aromatic and non-aromatic subgroups are further divided into binary subsets.

[0047] Each level stage employs a random forest binary classifier to randomly divide the input 17-dimensional spectral enhancement features into training and test sets, and performs standardization to eliminate scale differences.

[0048] After coarse classification at the first and second levels, a third-level random forest binary classifier performs fine classification in each chemical subset: the polypropylene (PP) subset and the polyethylene (PE) subset use a binary classifier to distinguish between polypropylene (PP) and polyethylene (PE); the polyvinyl chloride (PVC) subset directly outputs the PVC category; the aromatic subset's internal classifier further distinguishes between polystyrene (PS), polymethyl methacrylate (PMMA), and polyamide (PA); the non-aromatic subset's internal classifier distinguishes between polyethylene terephthalate (PET) and polycarbonate (PC).

[0049] Then, n-fold cross-validation was performed on all data in the training and test sets to obtain the average accuracy of the classification model. ;in The accuracy of the model when the nth data set is used as the validation set.

[0050] This innovative and improved model can transform complex multi-class classification problems into multiple chemically driven binary or subset classification tasks, thereby improving overall classification accuracy and achieving strong interpretability of the model.

[0051] The industrial-grade sorting device construction module is used to: construct a short-wave infrared multispectral imaging plastic sorting device that simulates an industrial production line using an InGaAs short-wave infrared detector, imaging lens, motorized filter wheel, miniature RGB camera and mechanical conveyor belt.

[0052] The motorized filter wheel has six circular slots, four of which contain bandpass filters for the four key wavelengths with the highest discrimination, while the other two slots are empty. The motorized filter wheel is used to transmit the full spectrum to achieve inter-frame positioning.

[0053] The miniature RGB camera monitors the conveyor belt in real time. When unsorted plastic is detected, the system starts the sorting process. Then, the electric filter wheel rotates at a predetermined speed, and the short-wave infrared multispectral camera takes pictures at the corresponding speed. The spectral data of the corresponding bands are collected through the bandpass filter on the electric filter wheel to obtain the spectral data of the four key bands with the highest discrimination. The inter-frame positioning is completed through the filterless slot.

[0054] After the computer performs the preprocessing operation b1 on the spectral data of the four key bands with the highest discrimination, it further calculates a 17-dimensional spectral enhancement feature vector dataset. Then, it calls the trained hierarchical decision tree (HDT) classification model for real-time inference and outputs the classification result image through hierarchical staged decision-making.

[0055] Compared with the prior art, the present invention has the following advantages:

[0056] (1) Significant cost advantage: By screening four key bands to construct a multispectral imaging system, the spectral bands are reduced in a reasonable and effective manner, thereby replacing the expensive hyperspectral camera, reducing the system hardware cost and complexity. At the same time, the short-wave infrared optical components can be manufactured using the same process as the visible light components, further reducing manufacturing costs and facilitating large-scale industrial application.

[0057] (2) High classification accuracy: The band screening method that integrates data-driven and materials chemistry knowledge effectively avoids the problems of high-dimensional data redundancy and feature loss. Combined with multi-dimensional feature extraction (4 basic features + 6 ratio features + 6 similarity features) and a hierarchical decision tree classification model driven by chemical properties, the classification accuracy is high and can accurately distinguish plastics of different materials, colors and shapes.

[0058] (3) Strong real-time dynamic sorting capability: The industrial-grade sorting device integrates RGB camera assistance, electric filter wheel switching band and inter-frame positioning, short-wave infrared multispectral camera for continuous data acquisition and upper computer calling classification model. The inference speed can reach 2fps / s, realizing real-time detection, real-time inference and continuous sorting of plastics, meeting the batch processing needs of industrial production lines.

[0059] (4) Good versatility: The band screening method and classification model of the present invention are not limited to specific types of plastics. They can be adapted to more plastic materials by expanding the spectral database. At the same time, the device design can be adjusted according to the needs of industrial production lines, and has strong flexibility and scalability.

[0060] (5) Strong environmental adaptability: The short-wave infrared band is not easily affected by Rayleigh scattering, can penetrate environments such as smoke and fog, and can shield color interference. Compared with visible light and long-wave infrared imaging, it can still maintain stable imaging quality and classification effect in complex industrial environments. Attached Figure Description

[0061] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0062] Figure 1 This is the overall flowchart of the method of the present invention.

[0063] Figure 2 This is a schematic diagram of a shortwave infrared hyperspectral data acquisition device.

[0064] Figure 3 This is a flowchart of the key feature band screening method in this invention.

[0065] Figure 4 This is a schematic diagram of a data acquisition device based on a shortwave infrared multispectral camera in key characteristic bands.

[0066] Figure 5 This is a schematic diagram of the cross-validation method.

[0067] Figure 6 This is the overall architecture diagram of the Hierarchical Decision Tree (HDT) classification model.

[0068] Figure 7 This is a schematic diagram of an imaging device for short-wave infrared multispectral data acquisition and plastic sorting.

[0069] Figure 8 This is a schematic diagram of the installation of the electric filter wheel.

[0070] Figure 9 This is a schematic diagram of an industrial-grade plastic sorting simulation platform device.

[0071] Figure 10 This is a pseudo-color image showing the multispectral classification results of eight plastic samples tested in practice. Detailed Implementation

[0072] Combination Figure 1 The present invention provides a low-cost short-wave infrared multispectral imaging plastic sorting system, including a spectral database establishment module, a key band screening module, a multispectral classification model construction module, and an industrial-grade sorting device construction module;

[0073] like Figure 2 As shown, the spectral database establishment module is used to perform the following steps:

[0074] Step a1: Install the shortwave infrared hyperspectral camera as the core acquisition device, install halogen lamps on both sides to provide illumination light in the 900~1700nm band, and then place the white calibration plate that provides the diffuse reflection reference and the black calibration plate that provides the noise reference on both sides of the electric push-broom stage.

[0075] Step a2: Place several plastic samples of the same type on an electric push-broom stage for each photograph to obtain spectral data of 224 bands in the range of 900~1700nm for each batch of samples, which is the raw brightness value DN (Digital Number) obtained by the camera sensor.

[0076] Step a3: Repeat step a2 to obtain spectral datasets for different types of plastics.

[0077] Combination Figure 3 The key band screening module is used to perform the following steps:

[0078] Step b1, perform the following preprocessing on the spectral dataset:

[0079] First, the outline of the plastic samples was marked using LabelMe software, and valid pixels within them were retained, resulting in a net dataset for each type of plastic sample. Then, the Interquartile Range (IQR) method was used to remove outliers, specifically removing outliers with DN values ​​exceeding a certain threshold. The above and The following data, of which , These are the first, third, and fourth quantiles of the dataset, respectively. .

[0080] Secondly, combining the spectral data from the white and black calibration plate regions, a normalization operation is performed on all the obtained valid data points. The specific formula is as follows: Where D is the new data after normalization. , , The data consist of valid data points, diffuse reflectance reference data within the white calibration plate area, and noise reference data within the black calibration plate area. After processing, the spectral reflectance dataset is obtained.

[0081] Step b2 involves reducing the dimensionality of the normalized spectral reflectance dataset obtained in the previous step to generate a pseudo-multispectral dataset.

[0082] Based on the wavelength-transmittance response curves of the candidate bandpass filter products, the data is integrated in segments according to different wavelength ranges, transforming the original 224-dimensional data into 42-dimensional pseudo-multispectral data. The specific formula is as follows:

[0083] ,

[0084] in, For wavelength The corresponding filter transmittance coefficient, For wavelength The nth dimension data (reflectivity value) corresponding to the band in question. wavelength infinitesimal increment, These are the lower and upper wavelength boundaries (i.e., the t-th dimension) corresponding to the t-th new segmented band interval, respectively. Let t be the spectral reflectance value of the new spectral dimension after one dimensionality reduction, according to We obtain a dimensionality-reduced spectral reflectance dataset.

[0085] Therefore, the original data is transformed into 42-dimensional pseudo-multispectral data, which reduces data redundancy while retaining key spectral information.

[0086] Step b3 involves a second dimensionality reduction, ultimately selecting four key bands.

[0087] This step is a feature extraction scheme that integrates data-driven approaches and materials chemistry knowledge. First, based on the dimensionality-reduced spectral reflectance dataset (42 dimensions) obtained in the previous step, a subset of bands is randomly selected as the initial feature subset ("seed" bands). On this subset of bands, the samples are classified using the K-means clustering method, and the ANOVA-F value of each attempt (reflecting the inter-group differences of the feature subset under different types of plastics) is calculated to verify the significance of the selected bands. The above process is repeated. Since the 42-dimensional dataset has a small dimension, a brute-force recursive method can be used to solve for the optimal band combination, and 6 candidate bands are selected.

[0088] Secondly, relevant literature was consulted, and the molecular structure and characteristic functional groups of various plastics under all candidate wavelengths were analyzed in conjunction with chemical property analysis. Simultaneously, differences were compared and verified, ultimately selecting four key wavelengths with the highest discriminative power: 1140nm, 1200nm, 1245nm, and 1310nm, for subsequent multispectral image classification. The reference results of the chemical properties of different types of plastics are shown in Table 1.

[0089] Table 1 Reference results for chemical properties

[0090]

[0091] Combination Figure 4 The multispectral classification model building module is used to collect data on different types of plastic samples using a short-wave infrared multispectral camera with an InGaAs detector as the core component, based on four selected key bands, to build a multispectral classification model and achieve accurate identification of plastic types. The specific steps are as follows:

[0092] Step c1: Fix the bandpass filters of the four selected key bands in front of the shortwave infrared multispectral camera (allowing only light of the corresponding wavelength to pass through), so that the camera, filters and plastic sample are on the same horizontal plane, and place the white calibration plate and black calibration plate on the left and right sides of the sample.

[0093] Step c2: Use a camera to take a horizontal picture of the plastic sample, while adjusting the height of the halogen lamp and the angle of the lampshade baffle so that the angle between the incident plane of the light and the horizontal plane in the image is... Values ​​in To ensure a stable and uniform short-wave infrared light source for the samples and avoid overexposure, the camera exposure time was adjusted according to the distance, and different types of plastic samples were photographed sequentially. Then, bandpass filters with different center wavelengths (bands) were used, and the above shooting steps were repeated (4 sets in total).

[0094] Step c3: Perform the same preprocessing operations (outlier removal and normalization) described above on all the obtained datasets.

[0095] Step c4: Using the reflectance dataset obtained in the previous step as the basic features, calculate the following various spectral enhancement features:

[0096] Original spectral eigenvectors ;in Spectral reflectance value under characteristic band i

[0097] inter-spectral ratio ;in , These are the spectral reflectance values ​​for characteristic bands i and j, respectively. The ratio between characteristic bands i and j is used here, resulting in a total of 6 sets of characteristic data.

[0098] Spectral similarity characteristics ; For the similarity features of characteristic bands i and j, a total of 6 sets of feature data were obtained here.

[0099] Finally, a 17-dimensional spectral enhancement feature vector dataset was obtained. ;in They are respectively The collection of all data Assign a number to the type of plastic it belongs to (different types of plastics are mapped to different numbers).

[0100] Step c5: Based on the 17-dimensional spectral enhancement feature vector dataset obtained in the previous step, an innovative hierarchical decision tree (HDT) classification model driven by chemical properties (HSE-Tree) is constructed to replace the traditional single-level decision tree (DT), such as... Figure 6 As shown, the specific construction process is as follows:

[0101] The first level separates polypropylene (PP) and polyethylene (PE) from other polymers. Utilizing the chemical properties of PP and PE as non-polar polycarbons, the characteristic absorptions produced by their C–H skeleton vibrations in the near-infrared spectrum show a clear spectral distinction from polymers containing polar groups or isomeric skeletons, thus achieving classification. This level distinguishes between non-polar polycarbons and polar polycarbons.

[0102] The second level: In non-PP and PE samples, polyvinyl chloride (PVC) is identified separately. The characteristic absorption differences caused by the chlorine-containing (C–Cl) chemical structure of PVC are used to achieve accurate differentiation. This level distinguishes chlorine-containing polymers from non-chlorine-containing polymers.

[0103] The third level: The remaining samples are divided into aromatic and non-aromatic polymers. Based on the combined vibrations and overtones of conjugated or polar groups such as aromatic rings, ester groups, and amide groups, a binary subset is formed. This level distinguishes between aromatic polymers and non-aromatic polymers.

[0104] Each of the above-mentioned stages uses a Random Forest binary classifier to randomly divide the input 17-dimensional spectral enhancement features into training and test sets, and performs standardization to eliminate scale differences.

[0105] After coarse classification at the first and second levels, fine classification is performed in each chemical subset using a third-level sub-classifier: a binary classifier is used within the PP / PE subset to distinguish between PP and PE; the PVC subset directly outputs the PVC category; the classifier within the aromatic subset further distinguishes between polystyrene (PS), polymethyl methacrylate (PMMA), and polyamide (PA); and the classifier within the non-aromatic subset distinguishes between polyethylene terephthalate (PET) and polycarbonate (PC).

[0106] After that, as Figure 5 As shown, n-fold cross-validation is performed on all data in the training and test sets to obtain the average accuracy of the classification model. ;in The accuracy of the model when the nth data set is used as the validation set. Figure 5 (vall_acc).

[0107] This innovative and improved model can transform complex multi-class classification problems into multiple chemically driven binary or subset classification tasks, thereby improving overall classification accuracy and achieving strong interpretability of the model.

[0108] Combination Figure 7 The industrial-grade sorting device module is used to construct a core device for data acquisition and plastic sorting using an InGaAs short-wave infrared detector, an imaging lens, and a motorized filter wheel. The motorized filter wheel has six circular slots; four slots contain bandpass filters for the selected key wavelengths, while the other two slots are unfiltered to allow full-spectrum transmission and inter-frame positioning. Figure 8 As shown.

[0109] Combination Figure 9 In addition to the components mentioned above, the entire sorting device constructed in this embodiment is also equipped with hardware systems such as halogen lamps, miniature RGB cameras, and mechanical conveyor belts. Among them, the RGB camera monitors the conveyor belt in real time. When it detects unsorted plastic passing through, the system starts the sorting process. Subsequently, the motorized filter wheel rotates at a speed of 2 r / s, and the short-wave infrared multispectral camera captures images at a speed of 12 fps / s. The spectral data of the corresponding wavebands are collected through the bandpass filters on the wheel, and inter-frame positioning is completed with the help of the filterless slot.

[0110] Subsequently, after performing the preprocessing operation b1 on the spectral data of the four key bands, the computer further calculates a 17-dimensional spectral enhancement feature vector dataset. Then, it calls the trained hierarchical decision tree (HDT) classification model for real-time inference and outputs the classification result image through hierarchical staged decision-making. When the system is working continuously, the output frame rate is maintained at about 2fps / s, which can realize continuous, real-time and accurate sorting of plastics on the conveyor belt.

[0111] The specific implementation method is as follows:

[0112] A batch of samples containing eight types of plastic fragments (PA, PET, PMMA, PP, PS, PVC, PC, and PE) was randomly divided into training and test sets in a 7:3 ratio. The 17-dimensional spectral enhancement feature vector data obtained from the training set samples after preprocessing was input into the hierarchical decision tree (HDT) classification model for training. Each random forest subclassifier in the model was set to 200 trees, and multi-core parallel computing was supported to improve training efficiency. In the cross-validation stage, the hierarchical prediction strategy of HDT was used to output the final category, and the model classification accuracy was 99.4%.

[0113] The test set samples are then processed through an industrial-grade sorting device and a short-wave infrared multispectral camera for real-time imaging and inference, outputting intuitive inference result images. For example... Figure 10 As shown, the spatial coordinates (x, y) and key spectral band values ​​of each plastic sample are reconstructed into a three-dimensional image cube. Three key bands (out of four) are arbitrarily selected and mapped to the red, green, and blue channels of pseudo-color, generating the original multispectral pseudo-color images of the eight plastic samples. Figure 10 The first and fifth rows (of the image) can intuitively reflect the differences in the distribution of different materials in the spectral space. After obtaining the pseudo-color image, the hierarchical decision tree (HDT) is called on the spectral features of each pixel to perform hierarchical classification prediction, generating the predicted category of each pixel, and mapping it to a color code to obtain the complete classification result image (the image is shown in the image). Figure 10 (Rows 2 and 6). Further, the classification results are overlaid on the pseudo-color image to form a semi-transparent overlay image (rows 2 and 6). Figure 10 The third and seventh rows of the map simultaneously calculate pixel-level classification accuracy. In the overlay map, color transparency reflects the correspondence between the classification result and the original spectral data; the clearer the target, the higher the accuracy. Furthermore, to visualize the correctness of the classification, a "correct / incorrect" mask map was designed. Figure 10 The 4th and 8th rows (indicated by the model) show the classification stability and potential error distribution of the model in different material regions. Green indicates correctly predicted pixels, and red indicates incorrectly predicted pixels. Actual testing showed a classification accuracy of 99.5% for the tested plastic samples.

[0114] In summary, this embodiment makes full use of the limited-band spectral data of plastics, and achieves low-cost, high-precision, and highly real-time multispectral plastic classification through mathematical feature enhancement and chemical structure-driven hierarchical classification. It is a simpler and lower-cost alternative to traditional hyperspectral plastic sorting methods, and provides a feasible technical path for resource recycling and reuse.

[0115] During this process, it is important to ensure that the light emitted by the halogen lamps can evenly cover the entire conveyor belt, and to place white calibration plates and black calibration plates on both sides of the conveyor belt.

[0116] The method and apparatus provided by this invention solve the problem of rapid, accurate, and low-cost sorting of different types of plastics on industrial production lines.

[0117] This invention provides a plastic sorting system based on short-wave infrared multispectral camera imaging. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A plastic sorting system based on short-wave infrared multispectral camera imaging, characterized in that, It includes a spectral database establishment module, a key band screening module, a multispectral classification model construction module, and an industrial-grade sorting device construction module; The spectral database establishment module is used to execute a method for rapidly acquiring short-wave infrared multi-band spectral data of different types of plastics; The key band selection module is used to perform a data preprocessing and fusion feature extraction method to select four key bands with the highest discriminative power from 224 original bands. The multispectral classification model building module constructs a multispectral classification model based on four key bands with the highest discriminative power to identify different types of plastics. The industrial-grade sorting device construction module is used to build a short-wave infrared multispectral imaging plastic sorting device that simulates an industrial production line, enabling real-time dynamic sorting.

2. The system according to claim 1, characterized in that, The method for rapidly acquiring short-wave infrared multi-band spectral data of different types of plastics includes the following steps: Step a1: Install the shortwave infrared hyperspectral camera as the core acquisition device, install halogen lamps on both sides of the shortwave infrared hyperspectral camera, the halogen lamps provide illumination light in the 900~1700nm band, and then place the white calibration plate that provides the diffuse reflection reference and the black calibration plate that provides the noise reference on both sides of the electric push-broom stage. Step a2: Place several plastic samples of the same type on an electric push-broom stage for imaging each time, and obtain spectral data of 224 bands in the range of 900~1700nm for each batch of samples, that is, the raw brightness value DN obtained by the internal sensor of the short-wave infrared hyperspectral camera. Step a3: Repeat step a2 to obtain spectral datasets for different types of plastics.

3. The system according to claim 2, characterized in that, The key band screening module is used to perform the following steps: Step b1: Preprocess the spectral dataset to obtain the spectral reflectance dataset; Step b2: Perform dimensionality reduction on the spectral reflectance dataset to generate a pseudo-multispectral dataset; Step b3 involves performing a second dimensionality reduction on the pseudo-multispectral dataset to ultimately select the four key bands with the highest discriminative power.

4. The system according to claim 3, characterized in that, Step b1 includes: marking the outline of the plastic sample and retaining the valid pixels inside, and organizing the data to obtain the net dataset for each type of plastic sample; then, using the interquartile range (IQR) method to remove outliers, i.e., removing DN values ​​in the range of 0.

05. The above and The following data, of which , These are the first, third, and fourth quantiles of the dataset, respectively. ; Combining the spectral data from the white and black calibration plates, a normalization operation is performed on all the obtained valid data points, using the following formula: Where D is the new data after normalization. , , The data are the valid data points, the diffuse reflectance reference data within the white calibration plate area, and the noise reference data within the black calibration plate area, respectively. The spectral reflectance dataset is obtained based on D.

5. The system according to claim 4, characterized in that, Step b2 includes: based on the band and transmittance response curves of the candidate bandpass filter products, performing integral calculations on the data segmented according to different band intervals, transforming the original data into 42-dimensional pseudo-multispectral data, as shown in the formula: ， in, For wavelength The corresponding filter transmittance coefficient, For wavelength The nth dimension data corresponding to the band in question. wavelength infinitesimal increment, These are the lower and upper wavelength boundaries corresponding to the t-th newly segmented band interval, respectively. Let t be the spectral reflectance value of the new spectral dimension after one dimensionality reduction, according to We obtain a dimensionality-reduced spectral reflectance dataset, where Z represents the set of integers.

6. The system according to claim 5, characterized in that, Step b3 includes: Step b3-1: Based on the dimensionality-reduced spectral reflectance dataset, randomly select a subset of bands as the initial feature set. Perform K-means clustering on the samples within this initial feature set to classify the features, and calculate the ANOVA-F value for each attempt to verify the significance of the selected bands. Step b3-2: Repeat step b3-1 and use a brute-force recursive method to solve for the optimal band combination, and select 6 candidate bands. Step b3-3: Combining chemical properties, analyze the molecular structure and characteristic functional groups of various types of plastics in all candidate wavelength bands, and conduct differential comparison verification. Finally, four key wavelength bands with the highest distinguishability are selected, namely 1140nm, 1200nm, 1245nm and 1310nm.

7. The system according to claim 6, characterized in that, The multispectral classification model construction module performs the following steps: Step c1: Select a short-wave infrared multispectral camera with an InGaAs detector as the core component. Fix the bandpass filters of the four key bands with the highest resolution in front of the short-wave infrared multispectral camera so that the camera, filters and plastic sample are on the same horizontal plane. At the same time, place the white calibration plate and the black calibration plate on the left and right sides of the sample. Step c2: Use a camera to take a horizontal picture of the plastic sample, while adjusting the height of the halogen lamp and the angle of the lampshade baffle so that the angle between the incident plane of the light and the horizontal plane is... Values ​​in Between these steps, the camera exposure time was adjusted according to the distance, and different types of plastic samples were photographed sequentially to obtain one set of data. Then, the bandpass filters with different center wavelengths were changed, and the photographing was repeated to obtain four sets of data in the end. Step c3: Perform the preprocessing of step b1 on the multispectral datasets of all types of plastic samples captured in step c2 under four different bandpass filters to obtain the normalized spectral reflectance dataset. Step c4: Using the normalized spectral reflectance dataset as the base feature, calculate the following spectral enhancement features: Original spectral eigenvectors ;in The spectral reflectance value under characteristic band i; inter-spectral ratio ;in , These are the spectral reflectance values ​​under characteristic band i and characteristic band j, respectively. The ratio between characteristic bands i and j was used to obtain a total of 6 sets of characteristic data. Spectral similarity characteristics ; Based on the similarity features of characteristic bands i and j, a total of 6 sets of feature data were obtained; Finally, a 17-dimensional spectral enhancement feature vector dataset was obtained. ;in They are respectively The collection of all data The collection of all data Number the type of plastic it belongs to; Step c5: Based on the 17-dimensional spectral enhanced feature vector dataset, construct a chemical property-driven hierarchical decision tree (HDT) classification model to replace the traditional single-level decision tree (DT).

8. The system according to claim 7, characterized in that, In step c5, the chemical property-driven hierarchical decision tree (HDT) classification model specifically includes: The first level: Polypropylene (PP) and polyethylene (PE) are separated from other polymers to obtain a subset of polypropylene (PP) and a subset of polyethylene (PE). Taking advantage of the chemical properties that both polypropylene (PP) and polyethylene (PE) are non-polar polyhydrocarbons, the characteristic absorption generated by their C-H skeleton vibration in the near-infrared spectrum has obvious spectral distinction with polymers containing polar groups or isomeric skeletons to achieve classification. The second level: In non-polypropylene (PP) and polyethylene (PE) samples, polyvinyl chloride (PVC) is identified separately to obtain a subset of PVC. The difference in characteristic absorption caused by the chlorination chemical structure of PVC is used to distinguish them. The third level: The remaining samples are divided into aromatic and non-aromatic polymers to obtain aromatic subsets and non-aromatic subsets. Based on the combined vibrations and overtones generated by the conjugation or polar groups of aromatic rings, ester groups, and amide groups, the aromatic and non-aromatic subgroups are further divided into binary subsets. Each level stage employs a random forest binary classifier to randomly divide the input 17-dimensional spectral enhancement features into training and test sets, and performs standardization to eliminate scale differences. After coarse classification at the first and second levels, a third-level random forest binary classifier performs fine classification in each chemical subset: the polypropylene (PP) subset and the polyethylene (PE) subset use a binary classifier to distinguish between polypropylene (PP) and polyethylene (PE); the polyvinyl chloride (PVC) subset directly outputs the PVC category; the aromatic subset's internal classifier further distinguishes between polystyrene (PS), polymethyl methacrylate (PMMA), and polyamide (PA); the non-aromatic subset's internal classifier distinguishes between polyethylene terephthalate (PET) and polycarbonate (PC). Perform n-fold cross-validation on all data in the training and test sets to obtain the average accuracy of the classification model. ;in The accuracy of the model when the nth data set is used as the validation set.

9. The system according to claim 8, characterized in that, The industrial-grade sorting device construction module is used to: construct a short-wave infrared multispectral imaging plastic sorting device that simulates an industrial production line using an InGaAs short-wave infrared detector, imaging lens, motorized filter wheel, miniature RGB camera and mechanical conveyor belt. The motorized filter wheel has six circular slots, four of which contain bandpass filters for the four key wavelengths with the highest discrimination, while the other two slots are empty. The motorized filter wheel is used to transmit the full spectrum to achieve inter-frame positioning. The miniature RGB camera monitors the conveyor belt in real time. When unsorted plastic is detected, the system starts the sorting process. Then, the electric filter wheel rotates at a predetermined speed, and the short-wave infrared multispectral camera takes pictures at the corresponding speed. The spectral data of the corresponding bands are collected through the bandpass filter on the electric filter wheel to obtain the spectral data of the four key bands with the highest discrimination. The inter-frame positioning is completed through the filterless slot.

10. The system according to claim 9, characterized in that, After the computer performs the preprocessing operation b1 on the spectral data of the four key bands with the highest discrimination, it further calculates a 17-dimensional spectral enhancement feature vector dataset. Then, it calls the trained hierarchical decision tree (HDT) classification model for real-time inference and outputs the classification result image through hierarchical staged decision-making.

Images