Machine vision-based material recycling analysis method and system

Abstract

The application provides a kind of material recycling analysis method and system based on machine vision, it is related to machine vision technical field, first, the visual acquisition data and acquisition ambient light parameters of the aluminum-plastic composite material to be recycled are acquired, and the superpixel unit cluster and optical response data are generated by superpixel segmentation processing;Then, based on the superpixel unit cluster and the optical response data corresponding to each superpixel unit, optical characteristic frequency domain analysis is performed, the frequency domain characteristic parameters specific to aluminum-plastic composite material are extracted, and a substance category determination model is constructed in combination with the frequency domain characteristic parameters;Call the substance category determination model to determine the category of the superpixel unit cluster and the optical response data of the aluminum-plastic composite material to be recycled to obtain the component partition result;According to the component partition result, an aluminum-plastic separation recycling instruction is generated and sent to a recycling sorting device, so that the components of the aluminum-plastic composite material can be accurately distinguished, efficient separation and recycling can be realized, the quality and purity of the recycled materials can be improved, and the recycling cost can be reduced.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and more specifically, to a material recycling analysis method and system based on machine vision. Background Technology

[0002] In the field of resource recycling, aluminum-plastic composites are a common packaging material widely used in the food and pharmaceutical industries. With increasing environmental awareness and growing demand for resource recycling, efficient and precise recycling of aluminum-plastic composites has become particularly important.

[0003] Currently, traditional methods for recycling aluminum-plastic composite materials mainly rely on manual sorting and simple mechanical separation. Manual sorting is not only inefficient but also susceptible to human error, resulting in low sorting accuracy and failing to meet the demands of large-scale recycling. Simple mechanical separation methods, such as density-based separation techniques, can achieve some separation of aluminum and plastic. However, the physical properties of aluminum and plastic in aluminum-plastic composites are not absolute, and the presence of contaminants such as stains and coatings on the material surface can lead to unsatisfactory separation results. Often, aluminum is mixed with plastic, or vice versa, reducing the quality and purity of the recycled material and consequently affecting its subsequent reuse value.

[0004] In addition, most existing recycling methods do not fully consider the characteristics of aluminum-plastic composite materials themselves, lack accurate identification and analysis of material components, and cannot formulate targeted separation and recycling strategies based on the actual composition of the materials, resulting in a lack of flexibility and adaptability in the recycling process. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a material recycling analysis method based on machine vision, the method comprising: The visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment are obtained. The visual acquisition data of the aluminum-plastic composite material to be recycled includes the surface texture details, color distribution differences and morphological structural features of the material. The illumination parameters of the acquisition environment include the light intensity variation data and the light incident angle information. The visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment are subjected to superpixel segmentation processing to generate superpixel unit clusters of the aluminum-plastic composite material to be recycled and optical response data corresponding to each superpixel unit. The optical response data reflects the reflection and refraction characteristics of the material under different illumination conditions. Based on the superpixel unit cluster and the optical response data corresponding to each superpixel unit, optical characteristic frequency domain analysis is performed to extract frequency domain feature parameters specific to aluminum-plastic composite materials. The material category determination model is constructed by combining the frequency domain feature parameters. The material category determination model is used to distinguish between aluminum component regions and plastic component regions in aluminum-plastic composite materials. The material category determination model is invoked to determine the category of the superpixel unit cluster and optical response data of the aluminum-plastic composite material to be recycled, and the composition partitioning result of the aluminum-plastic composite material to be recycled is obtained. The composition partitioning result includes the spatial location range of the aluminum component region and the spatial location range of the plastic component region. Based on the compositional zoning results of the aluminum-plastic composite material to be recycled, an aluminum-plastic separation and recycling instruction is generated. The aluminum-plastic separation and recycling instruction includes the sorting path planning of the aluminum component and the processing parameters of the plastic component, and the aluminum-plastic separation and recycling instruction is sent to the aluminum-plastic composite material recycling and sorting equipment.

[0006] Furthermore, embodiments of the present invention also provide a material recycling and analysis system based on machine vision, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described machine vision-based material recycling analysis method by executing the machine-executable instructions.

[0007] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described machine vision-based material recycling analysis method.

[0008] Based on the above, firstly, comprehensive visual acquisition data and ambient lighting parameters of the aluminum-plastic composite material to be recycled are obtained. This includes information such as surface texture details, color distribution differences, morphological features, light intensity variations, and incident angles. Then, superpixel segmentation processing is performed on the visual acquisition data and ambient lighting parameters to generate superpixel unit clusters and corresponding optical response data. This process characterizes the reflection and refraction properties of the material under different lighting conditions, dividing the complex surface features. Based on the superpixel unit clusters and optical response data, frequency domain analysis of optical properties is performed to extract frequency domain feature parameters specific to the aluminum-plastic composite material. A material classification model is then constructed, fully utilizing the material's optical properties to accurately distinguish between aluminum and plastic component regions. The material classification model is then used to determine the category, obtaining component partitioning results and clarifying the spatial location range of aluminum and plastic components. Based on the component partitioning results, an aluminum-plastic separation and recycling instruction is generated and sent to the recycling and sorting equipment. This achieves an automated process from analysis to execution, significantly improving recycling efficiency. This results in accurate analysis and efficient separation and recycling of aluminum-plastic composite materials, improving the quality and purity of the recycled materials and reducing recycling costs. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the material recycling analysis method based on machine vision provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of a machine vision-based material recycling and analysis system provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a schematic flowchart of a machine vision-based material recycling analysis method provided in one embodiment of the present invention. The following is a detailed description of the machine vision-based material recycling analysis method.

[0012] Step S110: Obtain visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment. The visual acquisition data of the aluminum-plastic composite material to be recycled includes material surface texture details, color distribution differences and morphological structural features. The illumination parameters of the acquisition environment include light intensity variation data and light incident angle information.

[0013] In this embodiment, the recycling and analysis of waste aluminum-plastic packaging materials is used as the application scenario. The aluminum-plastic composite material to be recycled in this scenario is mainly aluminum-plastic composite film used for food packaging. During the visual acquisition phase, an industrial camera is used to capture the aluminum-plastic composite material to be recycled from multiple angles. The camera resolution is set to a value suitable for capturing details on the material surface, and the lens focal length is adjusted according to the shooting distance to ensure that the surface texture of the material is clearly discernible in the image. During the shooting process, a light sensor is simultaneously activated to collect ambient light parameters. The sampling frequency of the light sensor is consistent with the camera's shooting frame rate to achieve time synchronization between the light data and the visual data. In the acquired visual acquisition data, the aluminum component areas typically exhibit a metallic luster and relatively uniform color distribution, while the plastic component areas show a specific plastic texture, and the color may vary depending on the added pigments. The morphological features include physical morphological information such as wrinkles and damage. Light intensity variation data is continuously acquired by the light sensor, reflecting the fluctuations in ambient light intensity over a period of time. The light incident angle information is obtained by an angle sensor installed next to the camera, recording the direction of light incident relative to the material surface.

[0014] Step S120: Perform superpixel segmentation processing on the visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment to generate a superpixel unit cluster of the aluminum-plastic composite material to be recycled and the optical response data corresponding to each superpixel unit.

[0015] In the aforementioned aluminum-plastic composite film recycling analysis scenario, superpixel segmentation is used to divide pixel regions with similar optical properties in an image into independent units, facilitating subsequent component analysis of each unit. Since the aluminum and plastic components in the aluminum-plastic composite film exhibit significant differences in optical properties, superpixel segmentation can effectively separate the regions containing these two components initially. The generation of superpixel unit clusters requires comprehensive consideration of the material surface texture, color, and lighting conditions, while optical response data is a quantitative description of the reflection, refraction, and other optical properties exhibited by each superpixel unit under different lighting conditions.

[0016] Step S121: Extract the edge contour information from the visual acquisition data of the aluminum-plastic composite material to be recycled, and enhance the pixel difference of the boundary transition area through the edge enhancement algorithm to obtain the edge-enhanced visual acquisition data. The edge contour information includes the boundary transition area features of the aluminum component and the plastic component.

[0017] In the visual acquisition data of aluminum-plastic composite films, a clear edge contour exists at the boundary between the aluminum and plastic components. The Canny edge detection algorithm is used to extract this edge contour information. It removes image noise through Gaussian filtering, calculates the image gradient, performs non-maximum suppression on the gradient magnitude, and finally determines the edge pixels using a double thresholding method. The extracted edge contour information includes not only the clear boundary between the aluminum and plastic components but also the internal edges formed by wrinkles, scratches, etc., on the material surface. To enhance the pixel differences in the boundary transition region, a Laplacian operator-based edge enhancement algorithm is used. This algorithm performs second-order differential operations on the image to highlight areas where grayscale values ​​change drastically, i.e., the boundary transition region. After edge enhancement processing, the pixel grayscale differences at the boundary between the aluminum and plastic components are more pronounced, enabling subsequent superpixel segmentation to more accurately identify and classify these two types of regions.

[0018] Step S122: Analyze the light intensity variation data in the light parameters of the acquired environment, determine the fluctuation period and amplitude of the light intensity, and set the time sampling interval for superpixel segmentation according to the fluctuation period of the light intensity so that the segmentation frequency matches the light fluctuation frequency.

[0019] At the aluminum-plastic composite film recycling and analysis site, light intensity may fluctuate periodically due to factors such as the switching of lighting equipment inside the workshop and changes in natural light outside the windows. Time-domain analysis is performed on the light intensity variation data collected by the light sensor. Fourier transform is used to convert the time-domain signal into a frequency-domain signal, and the main fluctuation frequencies are identified from the frequency-domain signal to calculate the light intensity fluctuation period. Simultaneously, the fluctuation amplitude is determined by calculating the difference between the maximum and minimum values ​​in the light intensity data. The time sampling interval for superpixel segmentation is set to half the light intensity fluctuation period. This is determined according to the sampling theorem to ensure accurate capture of the impact of light fluctuations on the optical properties of the material surface. This time sampling interval ensures that superpixel segmentation performs a sufficient number of samples within each cycle of light intensity change, thus enabling the segmentation results to reflect the changes in the material's optical properties under light fluctuations.

[0020] Step S123: Based on the incident light angle information in the illumination parameters of the collected environment, calculate the incident light projection angle of different areas on the surface of the aluminum-plastic composite material to be recycled, and set differentiated segmentation similarity thresholds for areas with different projection angles.

[0021] The surface of the aluminum-plastic composite film is not perfectly flat; it exhibits some curvature and undulations, resulting in varying incident light angles in different areas. Based on the incident light angle information and the three-dimensional coordinate data of the material surface (obtained through structured light scanning), the projection angle of the incident light at each point on the material surface is calculated using trigonometric principles. Specifically, the dot product of the incident light direction vector and the normal vector of each point on the material surface is performed, and then combined with inverse trigonometric functions to calculate the projection angle. For areas with significant differences in projection angles, their reflection characteristics also differ considerably; therefore, different segmentation similarity thresholds need to be set. For example, for areas with smaller projection angles, the light received by the material surface is more perpendicular, and the reflected light intensity is more stable; a higher segmentation similarity threshold can be set to avoid over-segmenting similar areas. Conversely, for areas with larger projection angles, the reflected light intensity is more sensitive to changes, and a lower segmentation similarity threshold is set to ensure that areas with subtle differences can be distinguished.

[0022] Step S1231: Obtain the three-dimensional morphological data of the aluminum-plastic composite material to be recycled. The three-dimensional morphological data includes the spatial coordinates and curvature information of each point on the material surface. The data is collected by a three-dimensional scanning device and converted into a standardized spatial coordinate matrix.

[0023] In the recycling analysis of aluminum-plastic composite films, a structured light 3D scanning device is used to scan the film. This device emits structured light of a specific pattern onto the material surface, and a camera captures the structured light pattern modulated on the surface. The triangulation principle is then used to calculate the 3D coordinates of each point on the material surface. During the scanning process, the scanning accuracy is adjusted according to the required detail of the material surface to ensure accurate acquisition of morphological features such as wrinkles and bends. The acquired 3D morphological data is transformed and unified into a preset world coordinate system, forming a standardized spatial coordinate matrix. Each element in this standardized spatial coordinate matrix represents the spatial coordinates of a point on the material surface. Furthermore, by calculating the coordinate data, the curvature information of each point can be obtained, reflecting the degree of bending of the material surface.

[0024] Step S1232: Based on the incident angle information in the illumination parameters of the acquired environment, determine the spatial direction vector of the incident light, and perform vector operation between the spatial direction vector of the incident light and the spatial coordinate matrix in the three-dimensional morphological data to obtain the incident light projection angle of different areas on the surface of the aluminum-plastic composite material to be recycled.

[0025] Given the incident light angle, it is converted into a spatial direction vector of the incident ray in the world coordinate system. For each point in the spatial coordinate matrix, the normal vector of that point is multiplied by the incident ray direction vector to obtain the cosine of the angle between them. The incident light projection angle is then calculated using the inverse cosine function. By traversing all points in the spatial coordinate matrix, the distribution of incident light projection angles in different regions of the material surface can be obtained.

[0026] Step S1233: Divide the surface of the aluminum-plastic composite material to be recycled into multiple angle intervals according to the incident light projection angle. Each angle interval contains a continuous projection angle range. Calculate the percentage of the material surface area contained in each angle interval.

[0027] Based on the calculated incident light projection angle, the angle range is divided into multiple continuous intervals, such as 0-15 degrees, 15-30 degrees, etc. For each angle interval, the number of points on the material surface where the projection angle falls within that interval is counted. Then, combined with the area weight corresponding to each point, the proportion of the material surface area contained in each angle interval is calculated. The area proportion information reflects the distribution of regions with different projection angles on the material surface.

[0028] Step S1234: Analyze the differences in optical reflection characteristics of aluminum-plastic composite materials in different angle ranges, determine the characteristic frequency range specific to the aluminum component and the plastic component through sample experiments, determine the differences in reflection characteristics of the aluminum and plastic components at different projection angles through sample experiments, and determine the preset reflection difference conditions.

[0029] Standard aluminum and plastic samples were selected, and different incident light projection angles were simulated in a laboratory environment. The reflectance data of the two samples at each angle were measured. Analysis of the reflectance data showed that the reflectance difference between the aluminum and plastic samples was significant in certain projection angle ranges, while the difference may be smaller in other angle ranges. Through statistical analysis of the above differences, a preset reflectance difference condition was determined. That is, when the reflectance difference between the two components in a certain projection angle range reaches a certain level, the reflectance characteristics difference in that angle range is considered to meet the preset condition.

[0030] Step S1235: Set differentiated segmentation similarity thresholds for regions with different projection angles. For regions where the reflection characteristics difference corresponding to the projection angle difference meets the preset reflection difference condition, use the segmentation similarity threshold that matches the reflection difference condition.

[0031] For projection angle regions that meet the preset reflection difference conditions, it indicates that the optical properties of aluminum and plastic components in this region are significantly different. To avoid misclassifying regions of different components as the same superpixel unit, a lower segmentation similarity threshold should be used to make it easier for the algorithm to classify them into different units. For regions whose reflection characteristics do not meet the preset conditions, a higher segmentation similarity threshold can be used to reduce the number of superpixel units and improve processing efficiency.

[0032] Step S1236: Map the spatial range of each angle interval to the pixel coordinates of the visual acquisition data of the aluminum-plastic composite material to be recycled, so that each pixel can correspond to its own angle interval and the corresponding segmentation similarity threshold.

[0033] By transforming coordinates, the spatial range of each angle interval in the three-dimensional morphological data is mapped to the pixel coordinates of the two-dimensional visual acquisition data. For each pixel in the visual image, its projection angle interval is determined based on its corresponding three-dimensional coordinates on the material surface, and then the segmentation similarity threshold corresponding to that projection angle interval is retrieved.

[0034] Step S1237: In the superpixel segmentation process, for each region to be segmented, first determine the angle interval to which the region to be segmented belongs, and call the segmentation similarity threshold corresponding to the angle interval for segmentation judgment, so that the segmentation process is adapted to the region projection angle.

[0035] During the execution of the superpixel segmentation algorithm, when processing a region in the image to be segmented, the algorithm first determines the projection angle interval to which the region belongs based on its pixel coordinates. Then, it retrieves the segmentation similarity threshold corresponding to that interval from the mapping relationship. When calculating the similarity of pixels within a region, the segmentation algorithm uses this threshold as a criterion. If the similarity between pixels is higher than the threshold, they are merged into the same superpixel unit; otherwise, they are divided into different units.

[0036] Step S1238: Record the usage of segmentation similarity thresholds and corresponding segmentation effects for different angle intervals. In subsequent analysis of similar material recycling, the segmentation similarity thresholds will be iteratively optimized based on historical segmentation effects.

[0037] After each superpixel segmentation, the segmentation similarity threshold used for each angular interval and the corresponding segmentation results are recorded, including segmentation performance indicators such as the number, size, and boundary accuracy of superpixel units. Through statistical analysis of the segmentation results of multiple batches of similar aluminum-plastic composite film materials, when a large number of missegmentations are found in the segmentation results of a certain angular interval, the segmentation similarity threshold for that interval is adjusted. For example, if aluminum and plastic regions are mistakenly merged, the threshold for that interval is lowered; if regions of the same component are over-segmented, the threshold is raised. This iterative optimization method continuously improves the accuracy of superpixel segmentation.

[0038] Step S124: Using a superpixel segmentation algorithm with a pre-set time sampling interval and differential similarity threshold, the edge-enhanced visual acquisition data is segmented to obtain an initial set of superpixel units.

[0039] The SLIC superpixel segmentation algorithm is applied to visually acquired data after edge enhancement of the aluminum-plastic composite film. The time sampling interval determined in step S122 is set as the iteration time step of the algorithm to ensure that the segmentation process is synchronized with the frequency of illumination fluctuations. Simultaneously, in the similarity calculation stage of the algorithm, the differential similarity threshold determined in step S123 is introduced. For pixels in different projection angle regions of the image, their corresponding thresholds are used for similarity judgment. The algorithm first initializes cluster centers, and then, within a specific region around each cluster center, pixels are assigned to the nearest cluster center based on color similarity and spatial distance similarity, forming initial superpixel units. After multiple iterations to optimize the cluster center positions until convergence, the initial set of superpixel units is finally obtained.

[0040] Step S125: Perform regional correlation analysis on the initial superpixel unit set, calculate the texture similarity and color similarity of adjacent superpixel units, merge adjacent superpixel units whose texture similarity and color similarity both exceed the set value, and remove isolated superpixel units whose area is smaller than the minimum effective area to form the superpixel unit cluster of the aluminum-plastic composite material to be recycled.

[0041] For each superpixel unit in the initial superpixel unit set, its texture and color features are extracted. Texture features are calculated using the gray-level co-occurrence matrix, including parameters such as energy, entropy, and contrast; color features are obtained using the average RGB color value of the superpixel unit. The texture similarity and color similarity between two adjacent superpixel units are calculated. Texture similarity is obtained by comparing the differences in their texture feature parameters, and color similarity is obtained by calculating the Euclidean distance between their average RGB color values. When the texture similarity and color similarity of two adjacent superpixel units both exceed a set threshold, they are merged into a new superpixel unit. Simultaneously, the area of ​​each superpixel unit is calculated, and isolated superpixel units with an area smaller than the minimum effective area (pre-set based on material size and analysis accuracy requirements) are removed. After merging and removing superpixel units, the final superpixel unit cluster is obtained.

[0042] Step S126: For each superpixel unit in the superpixel unit cluster, collect the pixel grayscale value change sequence of the superpixel unit under the current illumination intensity and incident angle, and calculate the reflectivity change curve and refractive index change curve of the superpixel unit through the grayscale value change sequence.

[0043] During the visual acquisition process of aluminum-plastic composite film, the grayscale value of each superpixel unit may change over time due to fluctuations in ambient light. To obtain the above-mentioned variation pattern, continuous grayscale value acquisition is performed on each superpixel unit. By converting the acquired grayscale value into reflectance, and then calculating the refractive index based on the relationship between reflectance and refractive index, a curve reflecting the change of the optical properties of the superpixel unit over time is obtained.

[0044] For example, in step S1261: during the visual acquisition process of the aluminum-plastic composite material to be recycled, a continuous acquisition mode is set, and multiple frames of visual images are acquired according to the preset acquisition interval time. The acquisition interval time is set to be less than the light intensity fluctuation period to capture the influence of light changes on grayscale values.

[0045] In the shooting parameter settings of the industrial camera, continuous shooting mode is enabled, and the acquisition interval is set according to the light intensity fluctuation period determined in step S122 to ensure that a sufficient number of image frames can be acquired within one fluctuation period. For example, if the light intensity fluctuation period is T, the acquisition interval is set to T / 10, so that 10 images can be acquired in each period, which can capture the influence of light intensity changes on the gray value of the material surface in a relatively detailed manner.

[0046] Step S1262: For each superpixel unit in the superpixel unit cluster, locate the pixel range of the superpixel unit in each frame of visual image, extract the grayscale values ​​of all pixels within the pixel range of the superpixel unit, and calculate the average value of the grayscale values ​​as the grayscale value of the superpixel unit at the current acquisition time.

[0047] For each frame of the visual image, the specific location of each superpixel unit in the image is determined based on the pixel coordinate range of each unit in the superpixel unit cluster. Then, all pixels within that pixel range are traversed, their grayscale values ​​are read, and the arithmetic mean of these grayscale values ​​is calculated. This arithmetic mean represents the overall grayscale level of the superpixel unit at the current acquisition time, reducing the impact of individual pixel noise on the results.

[0048] Step S1263: Arrange the gray values ​​of each superpixel unit at different acquisition times in chronological order to form a pixel gray value change sequence of the superpixel unit. Each element in the pixel gray value change sequence contains the acquisition time and the corresponding gray value.

[0049] The grayscale values ​​of the same superpixel unit in each frame are arranged sequentially according to the acquisition time, with each grayscale value corresponding to a specific acquisition time, thus forming a pixel grayscale value change sequence. This pixel grayscale value change sequence reflects the dynamic change of the grayscale value of the superpixel unit over time.

[0050] Step S1264: Obtain standard grayscale data of the acquisition environment. The standard grayscale contains multiple grayscale blocks with known reflectance. Under the same light intensity and incident angle, acquire visual images of the standard grayscale, extract the grayscale value of each grayscale block, and establish a correspondence model between grayscale value and reflectance. This correspondence model is constructed using linear fitting or polynomial fitting.

[0051] Under the same lighting and shooting conditions as the visual acquisition of the aluminum-plastic composite film, a standard grayscale plate was photographed. The standard grayscale plate contains multiple grayscale blocks with known reflectance values, which are calibrated at the factory. The grayscale values ​​of each block in the image were extracted. Using the grayscale value as the independent variable and the known reflectance as the dependent variable, a linear or polynomial fitting was performed using the least squares method to obtain a model of the correspondence between grayscale value and reflectance. For example, linear fitting can be expressed as reflectance equal to grayscale value multiplied by a coefficient plus a constant term, while polynomial fitting can construct a more complex relationship model based on the power terms of the grayscale value.

[0052] Step S1265: Input each gray value in the pixel gray value change sequence of the superpixel unit into the corresponding relationship model, calculate the reflectance of the superpixel unit at each acquisition time, arrange the reflectance in time order to form the reflectance change curve of the superpixel unit.

[0053] For each gray value in the sequence of gray value changes of a superpixel unit, it is substituted into the correspondence model established in step S1264, and the corresponding reflectance value is calculated through the model. Arranging these reflectance values ​​in chronological order of acquisition time yields the reflectance change curve of the superpixel unit, which visually demonstrates the trend of reflectance change of the superpixel unit over time.

[0054] Step S1266: Obtain the basic material parameters of the aluminum-plastic composite material to be recycled. The basic material parameters include the dielectric constant and magnetic permeability of the material.

[0055] By consulting the production data or relevant material manuals of aluminum-plastic composite films, we can obtain the basic material parameters such as the dielectric constant and magnetic permeability of the aluminum and plastic components. The dielectric constant and magnetic permeability of different materials are different, which leads to their different refractive indices.

[0056] Step S1267: According to electromagnetic theory, there is a definite mathematical relationship between reflectivity and refractive index. By substituting each reflectivity value in the reflectivity change curve into the formula based on Fresnel's law of reflection, the corresponding refractive index value is calculated.

[0057] Fresnel's law of reflection describes the reflection and refraction of light at the interface between two media. Based on this law, the relationship between reflectivity and refractive index can be derived. Given the surface reflectivity of a material, and combining this with parameters such as the material's permittivity, permeability, and the polarization state of the incident light, the refractive index of the material can be calculated using this relationship. For each reflectivity value in the reflectivity variation curve, the above calculation is performed sequentially to obtain the corresponding refractive index value.

[0058] Step S1268: During the calculation of refractive index, the air refractive index parameters of the collected environment are referenced. The air refractive index changes with temperature and humidity. The temperature and humidity data of the current environment are obtained through environmental monitoring equipment. The air refractive index parameters are corrected according to the formula of the influence of temperature and humidity on air refractive index.

[0059] The refractive index of air is slightly affected by ambient temperature and humidity, and these effects need to be considered in high-precision calculations. By collecting real-time temperature and humidity data from environmental monitoring equipment, and substituting this data into a known air refractive index correction formula, the accurate value of the air refractive index under the current environment can be calculated. This corrected air refractive index parameter is then used in the refractive index calculation process to improve the accuracy of the results.

[0060] Step S1269: Arrange the calculated refractive index values ​​in chronological order to form the refractive index change curve of the superpixel unit. The time axis of the refractive index change curve is consistent with that of the reflectivity change curve.

[0061] The calculated refractive index values ​​at each moment are arranged in chronological order of acquisition time to form a refractive index variation curve. This refractive index variation curve shares the same time axis as the reflectivity variation curve, facilitating comparative analysis of the optical property changes of the superpixel unit in terms of both reflectivity and refractive index.

[0062] Step S12610: Smooth the reflectivity change curve and the refractive index change curve. Use the moving average algorithm to remove random fluctuations in the curves and retain the overall trend. The size of the smoothing window is determined according to the degree of fluctuation of the curve.

[0063] Due to factors such as measurement noise, the reflectivity and refractive index variation curves may contain random fluctuations. A moving average algorithm is used to smooth the curves. An appropriate window size is selected, for example, a window containing 5 data points. The average value of the data within each window is calculated sequentially and used as the data value at the center of that window. This method effectively removes high-frequency noise from the curves while preserving the overall trend reflecting the changes in the material's optical properties.

[0064] Step S12611: Analyze the consistency between the reflectance change curve and the refractive index change curve, and calculate the correlation coefficient between the reflectance change curve and the refractive index change curve. A correlation coefficient close to 1 indicates that the change trends of the reflectance change curve and the refractive index change curve are consistent.

[0065] The Pearson correlation coefficient method was used to perform correlation analysis on the data of reflectance and refractive index variation curves. The correlation coefficient ranged from -1 to 1. When the correlation coefficient was close to 1, it indicated that the two curves showed a highly consistent trend, meaning that an increase or decrease in reflectance was accompanied by a corresponding increase or decrease in refractive index. When the correlation coefficient was close to -1, it indicated that the two curves showed opposite trends. A correlation close to 0 indicated a weak correlation. By analyzing the correlation coefficient, the consistency of the optical property changes reflected by the two curves could be verified.

[0066] Step S127: Integrate the reflectance change curve and refractive index change curve of each superpixel unit according to the time dimension, supplement the spatial position coordinate information corresponding to the superpixel unit, and form the optical response data corresponding to each superpixel unit.

[0067] For each superpixel unit, its reflectance and refractive index variation curves are aligned along the time dimension to ensure a one-to-one correspondence between the time points of the two curves. Simultaneously, the spatial coordinates of the superpixel unit in the visual acquisition data are added, such as the pixel coordinates of the unit's top-left and bottom-right corners. These data are then integrated to form a dataset containing time-series optical properties and spatial location information—that is, the optical response data corresponding to the superpixel unit.

[0068] Step S128: Establish a mapping index table between the superpixel unit cluster and the optical response data corresponding to each superpixel unit, and query the optical response data of any superpixel unit through the mapping index table.

[0069] To facilitate rapid retrieval of optical response data for each superpixel unit, a mapping index table is constructed. This index table uses the unique identifier of the superpixel unit as the key and the corresponding optical response data storage address as the value. When querying the optical response data of a specific superpixel unit, its unique identifier is used to search the index table, quickly locating the data's storage location and retrieving the data, thus improving data access efficiency.

[0070] Step S130: Based on the superpixel unit cluster and the optical response data corresponding to each superpixel unit, perform optical characteristic frequency domain analysis, extract frequency domain feature parameters specific to aluminum-plastic composite material, and construct a material category determination model in combination with the frequency domain feature parameters. The material category determination model is used to distinguish between aluminum component regions and plastic component regions in aluminum-plastic composite material.

[0071] In the scenario of aluminum-plastic composite film recycling analysis, the optical response characteristics of aluminum and plastic components differ in the frequency domain. By performing frequency domain analysis on the optical response data of each superpixel unit, frequency domain feature parameters that can characterize the differences between the two components are extracted. These parameters are then used to construct a material category determination model, thereby achieving accurate differentiation between aluminum and plastic component regions in aluminum-plastic composite materials. The model construction process includes data preprocessing, frequency domain transformation, feature extraction, dataset construction, and model training.

[0072] Step S131: Preprocess the optical response data corresponding to each superpixel unit in the superpixel unit cluster, and convert the preprocessed optical response data into a time series signal. The time series signal has the acquisition time as the horizontal axis and reflectivity or refractive index as the vertical axis. Each superpixel unit corresponds to a set of reflectivity time series signals and a set of refractive index time series signals.

[0073] For the optical response data of superpixel units, outlier detection and processing are first performed. The 3σ criterion is used to identify outliers in the reflectance and refractive index time series, and these outliers are replaced with the average value of the series or corrected using interpolation methods. Then, the time series is normalized, mapping the reflectance and refractive index data to the 0-1 range to eliminate the influence of dimensional differences on subsequent analysis. The normalized optical response data are plotted with time on the horizontal axis and reflectance or refractive index on the vertical axis, forming a time series signal. Each superpixel unit has two sets of time series signals: reflectance and refractive index.

[0074] Step S132: Perform a Fourier transform on the time series signal to convert the optical response change in the time domain into a spectral distribution in the frequency domain, and obtain the optical characteristic spectrum of each superpixel unit. The optical characteristic spectrum contains the amplitude and phase information of different frequency components.

[0075] The Fast Fourier Transform (FFT) algorithm is used to process the reflectivity and refractive index time-series signals, converting the time-domain signals into frequency-domain signals. In the resulting spectrum, the horizontal axis represents frequency, and the vertical axis represents the amplitude of each frequency component, also including phase information. The spectrum allows analysis of the various frequency components and their relative intensities within the optical response of the superpixel unit. Due to their different optical properties, aluminum and plastic components exhibit different frequency distribution characteristics in the spectrum.

[0076] Step S133: Analyze the spectral differences between the aluminum and plastic components in the aluminum-plastic composite material. Determine the characteristic frequency ranges specific to the aluminum component and the plastic component through sample experiments. Each characteristic frequency range contains at least two reliably distinguishable frequency peaks.

[0077] Pure aluminum and pure plastic samples were selected, and their optical characteristic spectrograms of superpixel units were obtained following the same acquisition and processing procedures as the aluminum-plastic composite material to be recycled. The spectrograms of the two types of samples were compared and analyzed to identify frequency regions with significant differences. Within these regions, the range containing at least two reliably distinguishable frequency peaks was selected as the characteristic frequency range. For example, aluminum samples exhibited obvious peaks at frequencies f1 and f2, while plastic samples showed lower or no peaks at these frequencies, but exhibited peaks unique to plastic samples at frequencies f3 and f4. Therefore, f1-f2 was determined as the characteristic frequency range specific to the aluminum component, and f3-f4 was determined as the characteristic frequency range specific to the plastic component.

[0078] Step S1331: Collect aluminum-plastic composite material samples with different material properties. The samples include aluminum-plastic composite materials with different aluminum content ratios and different types of plastics, such as polyethylene, polypropylene, and polyvinyl chloride, so that the samples cover common aluminum-plastic recycling scenarios.

[0079] Aluminum-plastic composite material samples were collected from various sources, including aluminum-plastic composite films used in common applications such as food packaging and pharmaceutical packaging. The aluminum content in the samples varied from low to high, and the plastic types covered common varieties such as polyethylene, polypropylene, and polyvinyl chloride. Ensuring sample diversity covered various material properties of aluminum-plastic composite materials that might be encountered during actual recycling, improving the versatility and accuracy of subsequent determination of characteristic frequency ranges.

[0080] Step S1332: Physically separate each aluminum-plastic composite material sample, extract the aluminum and plastic components from the sample to obtain pure aluminum and pure plastic samples.

[0081] The aluminum-plastic composite samples were separated using either chemical dissolution or mechanical exfoliation. For chemical dissolution, a specific solvent was used to dissolve the plastic component, leaving the aluminum component. For mechanical exfoliation, the difference in bonding strength between aluminum and plastic was utilized to separate them mechanically. The resulting pure aluminum and pure plastic samples were then cleaned to remove surface impurities and ensure sample purity.

[0082] Step S1333: Under the same acquisition environment and lighting conditions, acquire visual acquisition data and optical response data of pure aluminum samples and pure plastic samples respectively. Keep the acquisition device parameters consistent during the acquisition process. The acquisition device parameters are focal length, exposure time and sensitivity.

[0083] Pure aluminum and pure plastic samples were placed on the same test platform, using the same industrial camera and light sensor as those used when collecting data for the aluminum-plastic composite material to be recycled, and the same camera parameters such as focal length, exposure time, and sensitivity were set. Under the same ambient lighting conditions, visual data and lighting parameters of the two types of samples were collected, and their optical response data were calculated according to the method in step S126.

[0084] Step S1334: Following the same preprocessing procedure and Fourier transform parameters, perform frequency domain transformation on the optical response data of the pure aluminum sample and the pure plastic sample respectively to obtain the spectrum diagram of the pure aluminum sample and the spectrum diagram of the pure plastic sample.

[0085] The optical response data of pure aluminum and pure plastic samples are subjected to the same preprocessing and Fourier transform processing as in steps S131 and S132. The preprocessing includes outlier handling and normalization. The Fourier transform parameters, such as the number of sampling points and the type of window function, are kept consistent to ensure that the obtained spectrograms are comparable.

[0086] Step S1335: Compare the spectrograms of pure aluminum samples and pure plastic samples, find the regions of significant difference between the two in terms of spectral amplitude and frequency distribution, and record the frequency range corresponding to the regions of significant difference. The regions of significant difference are potential characteristic frequency ranges.

[0087] The spectra of pure aluminum and pure plastic samples were overlaid and compared to observe the differences in amplitude at different frequency positions. When there is a significant difference in the spectral amplitude between the two types of samples within a certain frequency range, and this difference is consistent across multiple samples, this frequency range is recorded as a potential characteristic frequency range.

[0088] Step S1336: For each potential characteristic frequency range, calculate the mean and standard deviation of the spectral parameters of multiple groups of pure aluminum samples and pure plastic samples within that frequency range, and analyze the degree of dispersion of the parameter distribution. The frequency range in which the degree of dispersion meets the preset dispersion condition and the difference between the mean parameters of the two groups of samples meets the preset difference condition is the candidate characteristic frequency range.

[0089] For each potential characteristic frequency range, spectral parameters, such as peak amplitude and total energy, are extracted from the spectrograms of multiple groups of pure aluminum and pure plastic samples. The mean and standard deviation of the parameters for each group are calculated; the standard deviation reflects the dispersion of the parameter distribution. A preset dispersion condition is that the standard deviation is less than a certain threshold to ensure a relatively concentrated distribution of sample parameters. A preset difference condition is that the difference between the means of parameters of two groups of samples is greater than a certain threshold to ensure that there is a significant difference in the parameters of the two types of samples within this frequency range. Frequency ranges that simultaneously meet both conditions are determined as candidate characteristic frequency ranges.

[0090] Step S1337: Verify the significance of the difference in spectral parameters between pure aluminum samples and pure plastic samples within the candidate characteristic frequency range using statistical testing methods. Set the significance level to a preset value, and determine the candidate characteristic frequency range that passes the significance test as the final characteristic frequency range.

[0091] Statistical tests such as t-tests or analysis of variance were used to test the significance of differences in spectral parameters between pure aluminum and pure plastic samples within the candidate characteristic frequency range. The significance level was preset to 0.05. A p-value less than 0.05 was considered statistically significant, and the test passed. The candidate characteristic frequency ranges that passed the test were determined as the final characteristic frequency range.

[0092] Step S1338: Divide the final characteristic frequency range into a characteristic frequency range specific to aluminum components and a characteristic frequency range specific to plastic components. The characteristic frequency range specific to aluminum components corresponds to a frequency range in which the spectral parameters of pure aluminum samples are significantly higher than those of pure plastic samples. The characteristic frequency range specific to plastic components corresponds to a frequency range in which the spectral parameters of pure plastic samples are significantly higher than those of pure aluminum samples.

[0093] Based on the relationship between the spectral parameters of pure aluminum and pure plastic samples within the final characteristic frequency range, the samples are divided into characteristic frequency ranges specific to aluminum and plastic components. When the mean spectral parameter of the pure aluminum sample is significantly higher than that of the pure plastic sample within a certain frequency range, that range is considered the characteristic frequency range specific to the aluminum component; conversely, it is considered the characteristic frequency range specific to the plastic component.

[0094] Step S1339: Verify the stability of the determined characteristic frequency range. Under different light intensities and different incident angles, repeatedly collect the spectral data of pure aluminum samples and pure plastic samples to check whether the characteristic frequency range remains stable. If there are fluctuations, adjust the frequency range boundary. Under different environments, distinguish the aluminum-plastic composition through this frequency range.

[0095] The light intensity of the test environment was varied, with multiple levels ranging from low to high illumination, and the incident angle was adjusted (e.g., 0 degrees, 30 degrees, 60 degrees, etc.). Spectral data of pure aluminum and pure plastic samples were collected under each condition. The effectiveness of the characteristic frequency range in distinguishing between the two types of samples under different conditions was analyzed. If the distinguishing effect of certain frequency ranges weakened with changes in environmental conditions, their boundaries were adjusted to ensure that the characteristic frequency range could stably distinguish between aluminum and plastic components under all possible environmental conditions.

[0096] Step S13310: Record the specific numerical ranges of the characteristic frequency ranges specific to the aluminum component and the plastic component, the corresponding spectral parameter characteristics, and the stability verification results to form a characteristic frequency range document.

[0097] The specific frequency ranges of the determined characteristic frequency ranges for aluminum and plastic components, the main spectral parameter characteristics within each range (such as the number of amplitude peaks, amplitude relationships, etc.), and the test results under different environmental conditions during the stability verification process are recorded in detail to form a characteristic frequency range document.

[0098] Step S134: Extract spectral parameters within the characteristic frequency range from the optical characteristic spectrum of each superpixel unit. The spectral parameters include the amplitude of the characteristic frequency peak, the bandwidth of the characteristic frequency peak, and the total energy value within the characteristic frequency range. The above spectral parameters constitute the frequency domain characteristic parameters specific to the aluminum-plastic composite material.

[0099] Based on the specific characteristic frequency ranges of the aluminum and plastic components determined in step S133, these frequency ranges are located in the optical characteristic spectrum of each superpixel unit. For each characteristic frequency range, the amplitude of the frequency peak (i.e., the ordinate value corresponding to the peak) is extracted; the bandwidth of the characteristic frequency peak (i.e., the frequency difference when the amplitude on both sides of the peak drops to half of the peak) and the total energy value within the characteristic frequency range are obtained by integrating the spectral amplitude within that range. The extracted spectral parameters are combined to form frequency domain characteristic parameters that can characterize the component to which the superpixel unit belongs.

[0100] Step S135: Collect superpixel unit clusters, optical response data and corresponding frequency domain feature parameters of different batches of aluminum-plastic composite materials to be recycled, label the actual component category of each superpixel unit as aluminum or plastic, and construct the training dataset for the material category determination model.

[0101] Multiple batches of aluminum-plastic composite materials from different sources were collected for recycling. Following the steps described above, superpixel segmentation, optical response data calculation, and frequency domain feature parameter extraction were performed. Then, through manual observation or other reliable component detection methods, the actual component category of each superpixel unit was labeled, categorizing it into aluminum or plastic. The frequency domain feature parameters of the superpixel units were used as input data, and the labeled categories were used as output labels to construct the training dataset for the material category determination model.

[0102] Step S136: Divide the training dataset into a training sample set and a validation sample set. The training sample set is used for model parameter learning, and the validation sample set is used for model performance verification. The division ratio is determined based on the total number of samples.

[0103] Based on the total number of samples in the training dataset, it is randomly divided into a training sample set and a validation sample set. When the total number of samples is large, a 7:3 or 8:2 ratio can be used, meaning 70% or 80% of the samples are used as the training sample set for model parameter learning; the remaining samples are used as the validation sample set to evaluate model performance during training and prevent overfitting. During the partitioning process, it is ensured that the proportion of aluminum and plastic samples in both the training and validation sample sets remains consistent with the original dataset.

[0104] Step S137: Construct the network architecture of the material category determination model. The network architecture includes an input layer, a feature fusion layer, a classification layer, and an output layer. The input layer receives the frequency domain feature parameters of the superpixel units. The feature fusion layer performs weighted fusion of different types of frequency domain feature parameters. The classification layer performs component category determination through a fully connected neural network. The output layer outputs the component category prediction results of the superpixel units.

[0105] The material classification model employs a deep learning network architecture. The number of neurons in the input layer matches the dimension of the frequency domain feature parameters, with each neuron corresponding to one frequency domain feature parameter. The feature fusion layer contains multiple neurons, each receiving different types of frequency domain feature parameter inputs. By setting different weight coefficients, these parameters are weighted and summed to achieve feature fusion. The fused features are then input to the classification layer, which consists of multiple fully connected neural network layers connected by activation functions. These layers perform nonlinear transformations and abstract representations of the input features. Finally, the output layer uses a softmax activation function to output the probability value of whether a superpixel unit belongs to aluminum or plastic. The category with the higher probability value is the model's prediction result.

[0106] Step S138: Input the training sample set into the network architecture, adjust the parameters of the material category determination model using the backpropagation algorithm, optimize the weights of the material category determination model by calculating the loss value between the prediction result and the labeled category, and use the cross-entropy loss function to calculate the loss value.

[0107] During model training, frequency domain feature parameters from the training sample set are input into the constructed network architecture. Forward propagation calculates the predicted probability of each sample's component class. Based on the predicted probability and the labeled class of the sample, the cross-entropy loss function is used to calculate the loss value, which effectively measures the difference between the predicted probability distribution and the true label distribution. Then, the backpropagation algorithm is used, starting from the output layer, to calculate the gradient of the loss value with respect to each network parameter (such as weights and biases) layer by layer. The parameter values ​​are adjusted according to the gradient direction and learning rate to reduce the loss value. Through multiple iterations of training, the model parameters are continuously optimized to improve the model's classification accuracy.

[0108] Step S1381: Initialize the network parameters of the material category determination model. The network parameters include the weight coefficients of the feature fusion layer, the weight matrix of the fully connected neural network of the classification layer, and the bias vector. The initialization adopts a random normal distribution to ensure that the initial parameter values ​​are within a reasonable range and to avoid gradient explosion or gradient vanishing in the early stage of training of the material category determination model.

[0109] The weight coefficients of the feature fusion layer, the weight matrix of the fully connected neural network in the classification layer, and the bias vector are initialized using a random normal distribution. For the weight coefficients and weight matrix, the mean is set to 0, and the standard deviation is set to a small value, such as 0.01, so that the initial parameter values ​​are distributed within a small range. The bias vector is initialized to 0 or a small positive value. This initialization method ensures that the output values ​​of each layer are reasonable in the early stages of model training, avoiding gradient explosion due to excessively large parameter values ​​or gradient vanishing due to excessively small parameter values, thus ensuring stable training.

[0110] Step S1382: Divide the training sample set into multiple training batches according to a preset batch size. Each training batch contains the same number of superpixel unit samples. The batch size is determined based on the memory capacity of the material category determination model training device.

[0111] Determine the appropriate batch size based on the GPU memory capacity or CPU memory size of the training device. Batch size represents the number of samples input into the model for parameter updates each time. When memory capacity is large, a larger batch size can be set, such as 64 or 128, to improve training efficiency; when memory capacity is small, a smaller batch size should be set, such as 16 or 32. Divide the training sample set into multiple training batches according to the batch size. If the last batch has insufficient samples, it can be handled by resampling or discarding.

[0112] Step S1383: Starting from the first training batch, input the sample data in the training batch into the input layer of the material category determination model. After processing by the feature fusion layer and the classification layer, the component category prediction probability of each sample is obtained. The component category prediction probability is the probability of aluminum component and the probability of plastic component.

[0113] The sample data from the first training batch, i.e., the frequency domain feature parameters of the superpixel units, are input into the input layer of the model. The input data is then passed to the feature fusion layer, where different frequency domain feature parameters are weighted and fused according to preset weight coefficients to obtain a fused feature vector. This fused feature vector is then input into the fully connected neural network of the classification layer. After linear transformations and activation function processing at each layer, the softmax function of the output layer calculates the predicted probability of each sample belonging to either aluminum or plastic components; the sum of these two probabilities is 1.

[0114] Step S1384: Based on the labeling category of the sample, which is aluminum or plastic, construct the true label vector. If the sample is labeled as aluminum, the true label vector is 1, 0; if the sample is labeled as plastic, the true label vector is 0, 1.

[0115] For each sample in the training batch, a true label vector is constructed based on its labeled component category. When a sample is labeled as aluminum, the true label vector is set to 1, 0; when labeled as plastic, the true label vector is set to 0, 1. This one-hot encoding method corresponds to the form of the predicted probability vector in the output layer, facilitating the calculation of the loss value.

[0116] Step S1385: Calculate the loss value between the predicted probability and the true label vector of each sample using the cross-entropy loss function. The cross-entropy loss function measures the difference between the two probability distributions.

[0117] For each sample, its component category predicted probability vector and true label vector are input into the cross-entropy loss function. The formula for calculating the cross-entropy loss function is: , where y_i is the i-th element of the true label vector and p_i is the i-th element of the predicted probability vector. This formula calculates the loss value for each sample; a larger loss value indicates a greater difference between the predicted probability distribution and the true label distribution.

[0118] Step S1386: Calculate the average loss value for each training batch. The average loss value is the arithmetic mean of the loss values ​​of all samples in that training batch. After obtaining the loss values ​​for all samples in the training batch, these loss values ​​are arithmetically averaged to obtain the average loss value for that training batch. The average loss value reflects the overall prediction error of the model on the current batch of samples.

[0119] Step S1387: Using the backpropagation algorithm, starting from the output layer of the material category determination model, calculate the partial derivative of the average loss value with respect to each network parameter layer by layer. Each network parameter is a weight coefficient, a weight matrix, and a bias vector. The partial derivative is the gradient value, which reflects the direction and degree of influence of parameter changes on the loss value.

[0120] The backpropagation algorithm starts from the output layer and calculates the partial derivatives of the average loss with respect to each parameter of the output layer, i.e., the gradient values, according to the chain rule. Then, these gradient values ​​are propagated back through the hidden layers (classification layer and feature fusion layer), calculating the gradient values ​​of the loss with respect to each hidden layer parameter in turn. The sign of the gradient value indicates the direction of the parameter change's influence on the loss value, and the magnitude of the gradient value indicates the degree of influence; the larger the gradient value, the more significant the impact of the parameter change on the loss value.

[0121] Step S1388: Based on the calculated gradient values, adjust the network parameters using the gradient descent optimization algorithm. The parameter adjustment method is that the new parameter value is equal to the old parameter value minus the learning rate multiplied by the gradient value, and the learning rate is set to a preset initial value.

[0122] The gradient descent optimization algorithm updates network parameters based on the calculated gradient values. For each network parameter, the new parameter value equals the old parameter value minus the product of the learning rate and the gradient value of that parameter. The learning rate is a preset initial value that controls the step size of parameter updates. An excessively large learning rate may lead to difficulty in model convergence, while an excessively small learning rate will slow down the training speed. The initial learning rate can be set to a small value, such as 0.001, and adjusted during training based on model performance.

[0123] Step S1389: After completing the parameter adjustment for one training batch, proceed to the next training batch and repeat the steps of sample input, prediction calculation, loss calculation, gradient calculation and parameter adjustment until all training batches have completed one training cycle. This process is considered one round of training.

[0124] After the parameters of the first training batch are adjusted, the next training batch is selected in sequence, and steps S1383 to S1388 are repeated, namely, inputting sample data, calculating the predicted probability, calculating the loss value, calculating the gradient value, and adjusting the parameters. When all training batches have completed the above training process once, one round of model training is completed.

[0125] Step S13810: After each round of training, the validation loss value and component category determination accuracy of the material category determination model are calculated using the validation sample set. The calculation method of the validation loss value and accuracy is the same as that of the training process.

[0126] After each round of training, the sample data from the validation sample set is input into the currently trained substance category determination model. The predicted probability of each sample is calculated in the same way as in the training process, and the validation loss value is calculated based on the predicted probability and the true label vector. At the same time, the category with the highest predicted probability is taken as the predicted category of the sample and compared with the true category. The proportion of correctly predicted samples to the total number of samples is counted to obtain the component category determination accuracy.

[0127] Step S13811: Compare the validation loss value and accuracy of the current round with those of the previous round. If the validation loss value decreases and the accuracy increases, continue to the next round of training. If the validation loss value increases for several consecutive rounds or the accuracy remains unchanged for several consecutive rounds, it is determined that the material category determination model training has reached convergence and training is stopped.

[0128] Compare the validation loss and accuracy of the current round with those of the previous round. If the validation loss decreases while the accuracy increases, it indicates that the model performance is continuously improving, and the next round of training is needed. If the validation loss shows an upward trend for several consecutive rounds (e.g., 3 or 5 rounds), or the accuracy remains unchanged for several consecutive rounds without further improvement, it indicates that the model training has reached convergence, and further training will not significantly improve the model performance; at this point, training should be stopped. Step S13812: During training, record the training loss, validation loss, and accuracy for each round of training, and plot the training curve. If overfitting occurs, with the training loss decreasing but the validation loss increasing, use dropout or L2 regularization to adjust the material category determination model.

[0129] During model training, the training loss, validation loss, and accuracy for each round are recorded in detail, and a training curve is plotted with the round number on the x-axis and these metrics on the y-axis. By observing the training curve, when the training loss continuously decreases while the validation loss begins to increase, it indicates that the model is overfitting, meaning that the model performs well on the training samples but has poor generalization ability. In this case, dropout technology can be used to randomly discard some neurons in the fully connected neural network of the classification layer to reduce co-fitting relationships between neurons; or L2 regularization can be used to add a squared term of the network parameters to the loss function to penalize excessively large parameter values, thereby suppressing overfitting.

[0130] Step S13813: After training stops, save the final material category determination model parameters and training process records. The final material category determination model parameters are the optimized weight coefficients, weight matrix, and bias vector.

[0131] Once the model training reaches convergence and stops, the optimized network parameters of each layer of the model, including the weight coefficients of the feature fusion layer, the weight matrix and bias vector of the fully connected neural network of the classification layer, as well as the training process records such as the training loss value, validation loss value and accuracy of each round, are saved to the specified storage location so that the model can be called later to determine the material category.

[0132] Step S139: Use the verification sample set to evaluate the performance of the material category determination model during the training process. When the evaluation index remains stable for multiple rounds of training and reaches the preset standard, stop the model training and obtain the final material category determination model. The evaluation index includes the accuracy of component category determination, the recall rate of aluminum component identification, and the recall rate of plastic component identification.

[0133] During model training, after each training round, the model's performance is evaluated using a validation sample set. The accuracy of component classification is the proportion of correctly predicted samples out of the total number of validation samples; the recall of aluminum component identification is the proportion of correctly predicted aluminum samples out of the actual number of aluminum samples in the validation sample set; and the recall of plastic component identification is the proportion of correctly predicted plastic samples out of the actual number of plastic samples in the validation sample set. Preset standard values ​​for these evaluation metrics are established, such as an accuracy of 90% or higher and a recall of 85% or higher. When these evaluation metrics remain stable across multiple training rounds and meet or exceed the preset standards, model training is stopped, resulting in the final material classification model.

[0134] Step S140: Call the material category determination model to determine the category of the superpixel unit cluster and optical response data of the aluminum-plastic composite material to be recycled, and obtain the composition partitioning result of the aluminum-plastic composite material to be recycled. The composition partitioning result includes the spatial location range of the aluminum component region and the spatial location range of the plastic component region.

[0135] In the scenario of aluminum-plastic composite film recycling analysis, the frequency domain feature parameters of each unit in the superpixel unit cluster of the aluminum-plastic composite material to be recycled are input into a trained material category determination model. The model outputs the component category prediction result for each superpixel unit. Based on the predicted category and spatial location information of each unit, adjacent units of the same type are merged to form continuous aluminum component regions and plastic component regions, and their spatial location range is determined, thereby obtaining the component partitioning result.

[0136] Step S141: The optical response data corresponding to each superpixel unit in the superpixel unit cluster of the aluminum-plastic composite material to be recycled is converted according to the input format required by the material category determination model. The converted optical response data includes a standardized representation of frequency domain feature parameters. The standardization process uses the standardized parameters used during the training of the material category determination model.

[0137] For each superpixel unit in the superpixel unit cluster of the aluminum-plastic composite material to be recycled, frequency domain feature parameters are extracted from its optical response data. Then, the frequency domain feature parameters are standardized according to the input data format requirements during the training of the material category determination model. During standardization, the same mean and standard deviation parameters as those used during model training are used to convert the frequency domain feature parameters into a standardized representation with a mean of 0 and a standard deviation of 1. This ensures that the distribution of the input data is consistent with the data distribution during model training, thereby improving the accuracy of model predictions.

[0138] Step S142: Input the converted optical response data into the input layer of the material category determination model, and perform dimensionality verification on the data to ensure that the dimension of the input data is consistent with the number of nodes in the input layer of the material category determination model. If the dimension does not match, return a data format error message and re-convert the data.

[0139] The standardized frequency domain feature parameters are input into the input layer of the material category determination model. During the data input process, the model first verifies the dimensionality of the input data, checking whether the number of features in the input data matches the number of neurons in the input layer. If the dimensionality does not match, the model will return a data format error message. At this point, the data transformation process needs to be re-examined to ensure that the number of extracted frequency domain feature parameters is correct, and the data transformation is repeated until the dimensionality of the input data matches the number of nodes in the model's input layer.

[0140] Step S143: Transmit the verified data to the feature fusion layer, and perform weighted summation on different types of frequency domain feature parameters according to preset weight coefficients. The different types of frequency domain feature parameters are the amplitude, bandwidth, and total energy value of the feature frequency peak, to obtain the fused feature vector. The weight coefficients are determined by the material category determination model training process and reflect the degree of contribution of different frequency domain features to the composition determination.

[0141] The validated input data is transmitted to the feature fusion layer of the material category determination model. In the feature fusion layer, frequency domain feature parameters (amplitude, bandwidth, and total energy value of characteristic frequency peaks) of different types are weighted and summed according to preset weight coefficients learned during model training. Each frequency domain feature parameter is multiplied by its corresponding weight coefficient, and all products are then summed to obtain a comprehensive fused feature vector. The magnitude of the weight coefficient reflects the contribution of the corresponding frequency domain feature parameter to the component determination; the larger the weight coefficient, the more important the feature parameter is in the component determination.

[0142] Step S144: Transmit the fused feature vector to the classification layer, and perform a nonlinear transformation on the fused feature vector through a fully connected neural network. Each neural network node processes the input data using a preset activation function, which is the ReLU function.

[0143] The fused feature vector is fed into the classification layer of the material category determination model, which consists of multiple fully connected neural networks. The fused feature vector is first input into the first layer of the classification layer, where each neuron multiplies each element of the fused feature vector with its corresponding weight matrix and adds a bias vector to obtain a linear combination result. This linear combination result is then processed by the ReLU activation function. The ReLU function sets input values ​​less than 0 to 0 and leaves input values ​​greater than 0 unchanged, thus achieving a non-linear transformation of the data and enhancing the model's expressive power. The data processed by the first layer of the neural network is then sequentially input into subsequent fully connected neural network layers, each undergoing similar linear transformations and ReLU activation function processing, continuously abstracting and extracting features.

[0144] Step S145: The classification layer outputs the probability value of each superpixel unit belonging to the aluminum component and the probability value of belonging to the plastic component. The probability value ranges from 0 to 1, and the sum of the two probability values ​​is 1.

[0145] After being processed by multiple layers of a fully connected neural network in the classification layer, the final feature vector is input to the output layer. The output layer uses the softmax activation function to process the input feature vector, converting it into a probability value for each superpixel unit belonging to either aluminum or plastic. The softmax function maps the elements of the feature vector to the range of 0 to 1, and the sum of all elements is 1, thus obtaining two probability values, representing the likelihood of the superpixel unit belonging to either aluminum or plastic.

[0146] Step S146: When the probability value of a superpixel unit belonging to aluminum is greater than the probability determination threshold, the superpixel unit is determined to be an aluminum component unit; when the probability value of a superpixel unit belonging to plastic is greater than the probability determination threshold, the superpixel unit is determined to be a plastic component unit; if neither probability value is greater than the probability determination threshold, the superpixel unit is marked as a unit to be confirmed.

[0147] A preset probability threshold, such as 0.5, is used. For each superpixel unit, the probability value of it belonging to the aluminum component is compared with this threshold. If the probability value is greater than the threshold, the superpixel unit is determined to be an aluminum component. Similarly, the probability value of it belonging to the plastic component is compared with the threshold. If the probability value is greater than the threshold, it is determined to be a plastic component. If both probability values ​​are less than or equal to the threshold, it indicates that the model has significant uncertainty in classifying the superpixel unit, and it is marked as a unit to be confirmed, requiring further processing.

[0148] Step S147: For superpixel units marked as units to be confirmed, extract the detailed features from the optical response data of the superpixel unit. The detailed features are the number of inflection points of the reflectance change curve and the fluctuation frequency of the refractive index. Supplement the input to the classification layer of the material category determination model and recalculate and determine the probability.

[0149] For superpixel units marked as unconfirmed units, detailed features are re-extracted from their optical response data. These include the number of inflection points in the reflectance variation curve (determined by finding the point where the derivative is zero), and the refractive index fluctuation frequency (obtained through spectral analysis of the refractive index variation curve). These detailed features are then added to the existing frequency domain feature parameters to form a new feature vector, which is then input again into the classification layer of the material category determination model for recalculation of probabilities and re-determination of the category. By adding detailed features, the model's accuracy in identifying uncertain units is improved.

[0150] Step S148: Collect all superpixel units that are determined to be aluminum components, analyze the spatial distribution of the superpixel units, merge adjacent aluminum components to form a continuous aluminum component region, and record the boundary coordinates and area size of each aluminum component region.

[0151] All superpixel units identified as aluminum-based are collected, and their spatial distribution is analyzed based on their spatial coordinates in the visual acquisition data. Spatially adjacent aluminum-based units are merged into a single, continuous aluminum-based region. The boundary coordinates (i.e., the pixel coordinates of the top-left and bottom-right corners) of the merged region are determined by calculating its minimum bounding rectangle. Simultaneously, the number of superpixel units within each region is counted, and the area of ​​each aluminum-based region is calculated based on the area of ​​each unit.

[0152] Step S149: Using the same method, merge adjacent plastic component units to form a continuous plastic component region, and record the boundary coordinates and area size of each plastic component region.

[0153] Following the same method as in step S148, superpixel units identified as plastic component units are processed. All plastic component units are collected, their spatial distribution is analyzed, adjacent plastic component units are merged to form continuous plastic component regions, and the boundary coordinates and area size of each region are recorded.

[0154] Step S1410: Integrate the information of all aluminum and plastic component regions, including the number of regions, the boundary coordinates of each region, the area size, and the corresponding number of superpixel units, to form the component partitioning result of the aluminum-plastic composite material to be recycled.

[0155] Information on all aluminum and plastic component regions is integrated, including the total number of regions, the boundary coordinates and area of ​​each region, and the number of superpixel units contained in that region. This information is then organized into a structured data format to form the composition partitioning result of the aluminum-plastic composite material to be recycled. This composition partitioning result shows the spatial distribution of aluminum and plastic components in the material.

[0156] Step S1411: Convert the component partitioning results into a visualization image, and overlay the boundary markers of the aluminum component area and the plastic component area on the original visual acquisition data, using different colors to distinguish different component areas.

[0157] To visually represent the composition zoning results, they were converted into a visual image. On the original visually acquired image of the aluminum-plastic composite material to be recycled, boundary markers for the aluminum and plastic component areas were drawn based on the boundary coordinates of each region in the composition zoning results. For the aluminum component areas, a specific color (such as red) was used for filling or marking the boundary lines; for the plastic component areas, another specific color (such as blue) was used for marking. This method clearly distinguishes the areas containing the aluminum and plastic components on the image.

[0158] Step S150: Based on the composition zoning results of the aluminum-plastic composite material to be recycled, generate an aluminum-plastic separation and recycling instruction. The aluminum-plastic separation and recycling instruction includes the sorting path planning of the aluminum component and the processing parameters of the plastic component, and send the aluminum-plastic separation and recycling instruction to the aluminum-plastic composite material recycling and sorting equipment.

[0159] In the scenario of aluminum-plastic composite film recycling analysis, based on the spatial location and area of ​​the aluminum and plastic component regions in the component zoning results, a sorting path plan for the aluminum component is formulated, determining the sequence and path for the robotic arm to grasp the aluminum region. Simultaneously, based on the characteristic parameters of the plastic component, processing parameters such as heating treatment temperature, separation process parameters, and subsequent processing parameters are determined. These plans and parameters are integrated into an aluminum-plastic separation and recycling instruction, which is sent to the recycling and sorting equipment according to its communication protocol, guiding the equipment to perform the aluminum-plastic separation and recycling operation.

[0160] Step S151: Analyze the component zoning results of the aluminum-plastic composite material to be recycled, extract the spatial location range of the aluminum component area and the spatial location range of the plastic component area, and convert the spatial location range into coordinate data consistent with the coordinate system of the recycling and sorting equipment, so that the recycling and sorting equipment can accurately locate the component area.

[0161] The composition zoning results are analyzed to extract spatial location information such as the boundary coordinates of the aluminum and plastic component regions. Since the coordinates in the composition zoning results are based on the image coordinate system of the visually acquired data, while recycling and sorting equipment typically has its own equipment coordinate system, coordinate transformation is required. By establishing a transformation relationship between the image coordinate system and the equipment coordinate system, such as through a transformation matrix obtained from calibration, the spatial location range of the component regions is converted from the image coordinate system to coordinate data in the equipment coordinate system. This ensures that the recycling and sorting equipment can accurately locate the aluminum and plastic component regions based on this coordinate data.

[0162] Step S152: Analyze the spatial distribution characteristics of the aluminum component regions, and classify the priority of aluminum components according to the area size of the aluminum component regions. The spatial distribution characteristics include the number of aluminum component regions, the area size of each aluminum component region, and the relative positional relationship between each aluminum component region. The aluminum component regions whose area size meets the preset area conditions are set as the sorting priority matching the area conditions.

[0163] The spatial distribution characteristics of aluminum composition regions were analyzed, the number of aluminum composition regions was counted, the area of ​​each region was measured, and the relative distance between regions was calculated. Based on preset area criteria, regions with areas larger than a certain threshold were classified as high priority, regions with areas smaller than that threshold but larger than another threshold were classified as medium priority, and the rest as low priority. The aluminum composition regions were then sorted according to priority; larger regions generally had higher sorting priority to prioritize the processing of higher-value or more abundant aluminum components.

[0164] Step S153: Based on the coordinate data and priority of the aluminum composition region, and combined with the movement range, movement speed and sorting accuracy parameters of the robotic arm of the recycling and sorting equipment, plan the sorting path of the aluminum composition.

[0165] Based on the coordinate data and priority of the aluminum composition regions in the equipment coordinate system, and combined with the motion parameters of the robotic arm of the recycling and sorting equipment, an efficient sorting path is planned. The range of motion of the robotic arm determines the spatial area it can reach, while the motion speed and sorting accuracy parameters affect the sorting efficiency and accuracy. Path planning needs to ensure that the robotic arm can reach each aluminum composition region sequentially, with the shortest possible path, avoiding unnecessary back-and-forth movements, while also meeting the sorting accuracy requirements.

[0166] Step S1531: Obtain the motion parameters of the robotic arm of the recycling and sorting equipment. The motion parameters of the robotic arm include the maximum motion angle of each joint of the robotic arm, the upper limit of motion speed, the upper limit of motion acceleration, and the positioning accuracy error.

[0167] The motion parameters of the robotic arm can be obtained by consulting the technical manual of the recycling and sorting equipment or communicating with the equipment control system. The maximum motion angle of each joint of the robotic arm limits the rotation range of the joint, the upper limit of motion speed and acceleration determines the speed and acceleration / deceleration capability of the robotic arm, and the positioning accuracy error represents the allowable deviation range between the actual position reached by the robotic arm and the target position. The above parameters are the constraints for sorting path planning.

[0168] Step S1532: Convert the coordinate data of the aluminum composition region into target coordinate points in the robotic arm coordinate system. Each aluminum composition region corresponds to at least one target coordinate point. The target coordinate point is selected as the geometric center of the aluminum composition region, so that the robotic arm can accurately grasp the aluminum composition region.

[0169] Based on the transformation relationship between the robotic arm coordinate system and the equipment coordinate system, the coordinate data of the aluminum composition region is converted into coordinates in the robotic arm coordinate system. For each aluminum composition region, its geometric center is selected as the target coordinate point. The geometric center is obtained by calculating the average value of the region boundary coordinates. Selecting the geometric center as the target coordinate point enables the robotic arm to more accurately locate the center of the aluminum composition region during grasping, thereby improving the grasping success rate.

[0170] Step S1533: Sort the target coordinate points according to the priority of the aluminum composition area. The target coordinate points corresponding to the aluminum composition area that meets the preset priority conditions are placed at the front of the sorting sequence. At the same time, the sorting is adjusted based on the priority, taking into account the spatial distance between the target coordinate points, to reduce the round-trip movement distance of the robotic arm.

[0171] First, the target coordinates are initially sorted according to the priority of the aluminum composition regions, with higher-priority regions' target coordinates placed at the front of the sorting sequence. Then, based on the initial sorting, the spatial distance between adjacent target coordinates is considered, and a greedy algorithm or other path optimization algorithm is used to adjust the sorting. For example, when priorities are the same, spatially adjacent target coordinates are arranged in a consecutive sorting sequence to reduce the round-trip distance of the robotic arm between different target points and improve sorting efficiency.

[0172] Step S1534: Calculate the straight-line distance between two adjacent target coordinate points. Combine the upper limit of the robotic arm's movement speed and the upper limit of its movement acceleration to estimate the theoretical movement time for the robotic arm to move from one target coordinate point to the next. The theoretical movement time includes the sum of the acceleration phase, the constant speed phase, and the deceleration phase.

[0173] For the sorted sequence of target coordinate points, the straight-line distance between any two adjacent target coordinate points is calculated sequentially. Based on the upper limits of the robotic arm's speed and acceleration, kinematic formulas are used to estimate the time it takes for the robotic arm to travel between two target points. The robotic arm's motion typically includes an acceleration phase, a constant speed phase, and a deceleration phase. In the acceleration phase, the robotic arm accelerates from rest to its maximum speed limit at maximum acceleration. If the distance is long enough, it enters the constant speed phase, moving at its maximum speed. Finally, in the deceleration phase, it decelerates to its maximum speed and stops at the target point. By calculating and summing the times for each of these three phases, the theoretical motion time is obtained.

[0174] Step S1535: Based on the positioning accuracy error of the robotic arm, a safe operating range is set around each target coordinate point. The size of the safe operating range enables the robotic arm to grasp the aluminum component area even when there is a positioning error.

[0175] Considering the positioning accuracy error of the robotic arm, a safe operating range is set around each target coordinate point. The size of the safe operating range is usually set as a multiple of the positioning accuracy error (e.g., 2-3 times), and adjusted according to the size of the aluminum component area. This ensures that even if there is a certain positioning deviation when the robotic arm moves to the target coordinate point, its end effector can still fall within the safe operating range, thereby accurately grasping the aluminum component area.

[0176] Step S1536: Plan the initial sorting path. The initial sorting path is generated by connecting adjacent target coordinate points with straight lines according to the sorted target coordinate points, while avoiding obstacles within the movement range of the robotic arm. The obstacles are the equipment frame and other materials. The location information of the obstacles is obtained from the environmental map of the recycling and sorting equipment.

[0177] Based on the adjusted target coordinates, an initial sorting path is generated by connecting adjacent target coordinates with straight lines. During path generation, obstacle detection and avoidance are required. Obstacle location information, such as the coordinate range of the equipment rack and the location of other materials to be processed, is obtained from the environmental map of the recycling and sorting equipment. A collision detection algorithm is used to check whether the initial path collides with obstacles. If a collision occurs, the path is adjusted, such as by taking a detour to avoid the obstacle, generating an initial sorting path that can safely reach each target coordinate point.

[0178] Step S1537: Optimize the initial sorting path, use a path smoothing algorithm to adjust the path curve, and calculate the total length and total movement time of the optimized sorting path. The total movement time meets the industrial-grade detection throughput requirements, and the number of aluminum-plastic composite materials processed per unit time is not less than the preset throughput standard. If the total movement time exceeds the preset time range, the target coordinate point sorting is readjusted or the movement speed parameters are optimized.

[0179] The initial sorting path is smoothed using B-spline curves or other path smoothing algorithms to eliminate sharp corners and abrupt changes, resulting in smoother robotic arm movement and reduced impact and vibration. The total length of the optimized sorting path is then calculated, and the total movement time is estimated based on the robotic arm's speed parameters. This total movement time is compared to a preset time range, determined according to industrial-grade inspection throughput requirements, ensuring that the number of aluminum-plastic composite materials processed per unit time is not less than the preset standard. If the total movement time exceeds the preset range, the sorting order of the target coordinate points needs to be readjusted, or the robotic arm's speed parameters need to be appropriately increased (within the equipment's allowable range), and the path replanned and optimized until the total movement time meets the requirements.

[0180] Step S1538: Record the coordinate sequence, motion speed curve and motion time allocation of the final determined sorting path, and integrate them into sorting path planning data for aluminum composition.

[0181] The optimized sorting path target coordinate point sequence, the robotic arm's motion speed curves on each path segment (including speed changes during acceleration, constant speed, and deceleration phases), and the motion time allocation for each path segment are recorded in detail. This data is then integrated into structured aluminum composition sorting path planning data, which serves as part of the aluminum-plastic separation and recycling instruction.

[0182] Step S154: Extract material characteristic information of the plastic component region. The material characteristic information is derived from the frequency domain feature parameters of the material category determination model. The material characteristic information includes the melting temperature range, density parameter and chemical stability characteristics of the plastic component. The melting temperature range, density parameter and chemical stability characteristics of the plastic component are used to determine the processing parameters of the plastic component and to determine the processing technology of the plastic component.

[0183] Based on the frequency domain characteristic parameter analysis results of the superpixel units in the plastic composition region using the material category determination model, and combined with the known spectral characteristic database of plastic materials, the material properties of the plastic composition are derived. Different types of plastics have different melting temperature ranges, densities, and chemical stability. For example, polyethylene and polypropylene have different melting temperatures, requiring different heating temperatures for processing; density differences can be used to separate plastics from other impurities; and chemical stability characteristics determine the processing conditions that plastics can withstand.

[0184] Step S155: Determine the heat treatment temperature parameters of the plastic component based on its melting temperature range. Control the heat treatment temperature within the melting temperature range to ensure effective melting of the plastic component without excessive degradation.

[0185] Based on the derived melting temperature range of the plastic components, the heat treatment temperature parameters are set in the lower-middle range of this range. For example, if the melting temperature range is T1-T2, the heat treatment temperature can be set to approximately (T1+T2) / 2 to ensure that the plastic components can melt effectively, facilitating subsequent processing. Simultaneously, controlling the temperature within the melting temperature range prevents excessive degradation of the plastic components due to excessively high temperatures, which would affect their recycling value.

[0186] Step S156: Based on the density parameters of the plastic components, a corresponding separation process is adopted. The separation process is density sorting. The density parameters of the sorting medium are determined so that the plastic components are separated from other impurities through density differences.

[0187] Based on the density parameters of the plastic components, density sorting was selected as the separation process. Density sorting utilizes the density differences between different substances to achieve separation in a sorting medium of a specific density. The density parameters of the sorting medium were determined to be between the density of the plastic components and the density of other impurities. When the plastic components and impurities are placed in the sorting medium, the plastic components with a density less than that of the sorting medium will float to the surface, while the impurities with a density greater than that of the sorting medium will sink, thus achieving the separation of the plastic components from the impurities.

[0188] Step S157: Based on the chemical stability characteristics of the plastic components, determine the process parameters for subsequent processing. The subsequent processing is granulation and injection molding. The process parameters include processing pressure, processing time, and cooling rate, so that the plastic components are processed into qualified recycled raw materials.

[0189] Based on the chemical stability characteristics of the plastic components, appropriate subsequent processing techniques, such as granulation or injection molding, should be selected. For plastic components with good chemical stability, higher processing pressure and longer processing time can be used; for plastic components with poor chemical stability, processing pressure should be reduced and processing time shortened to avoid material decomposition. Simultaneously, based on the thermal properties of the plastic, an appropriate cooling rate should be determined to ensure that the processed plastic product has good physical properties, thus transforming the plastic components into qualified recycled raw materials.

[0190] Step S158: Integrate the sorting path planning data of aluminum components and the processing parameters of plastic components, and generate an aluminum-plastic separation and recycling instruction according to the instruction format requirements of the recycling and sorting equipment. The aluminum-plastic separation and recycling instruction includes the instruction generation time, the identifier of the material to be processed, and the equipment operation sequence.

[0191] The sorting path planning data for aluminum components and the processing parameters for plastic components are integrated. The data is then organized into a specific instruction structure according to the instruction format requirements that the recycling and sorting equipment can recognize. In addition to the sorting path planning data and processing parameters, the aluminum-plastic separation and recycling instruction must also include the instruction generation time for equipment recording and traceability; a material identification to identify the batch or individual aluminum-plastic composite material to be processed; and a detailed equipment operation sequence, clearly defining the action order and parameter settings of each component.

[0192] Step S159: Send the aluminum-plastic separation and recycling instruction to the aluminum-plastic composite material recycling and sorting equipment.

[0193] The generated aluminum-plastic separation and recycling instructions are sent to the control system of the aluminum-plastic composite material recycling and sorting equipment via industrial Ethernet or other communication methods. Upon receiving the instructions, the equipment control system parses the instructions and, according to the sorting path planning and processing parameters in the instructions, controls the robotic arm and other processing mechanisms to perform the aluminum-plastic separation and recycling operation. During the instruction transmission process, a data verification mechanism is employed to ensure the integrity and accuracy of the instruction data. If transmission fails, it is retransmitted until the equipment successfully receives the instructions.

[0194] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a machine vision-based material recycling analysis system 100 provided in this application embodiment for performing the above-described machine vision-based material recycling analysis method. The machine vision-based material recycling analysis system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0195] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located in the machine vision-based material recycling analysis system 100 and are separately configured. Alternatively, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the machine vision-based material recycling analysis method provided in the aforementioned method embodiments.

[0196] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A material recycling analysis method based on machine vision, characterized in that, The method includes: The visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment are obtained. The visual acquisition data of the aluminum-plastic composite material to be recycled includes the surface texture details, color distribution differences and morphological structural features of the material. The illumination parameters of the acquisition environment include the light intensity variation data and the light incident angle information. The visual acquisition data of the aluminum-plastic composite material to be recycled and the illumination parameters of the acquisition environment are subjected to superpixel segmentation processing to generate a superpixel unit cluster of the aluminum-plastic composite material to be recycled and optical response data corresponding to each superpixel unit. This includes: extracting edge contour information from the visual acquisition data of the aluminum-plastic composite material to be recycled, strengthening the pixel difference of the boundary transition region through an edge enhancement algorithm, and obtaining edge-enhanced visual acquisition data. The edge contour information includes the boundary transition region features of the aluminum component and the plastic component. Analyze the light intensity variation data in the light parameters of the acquired environment to determine the fluctuation period and amplitude of the light intensity. Set the time sampling interval for superpixel segmentation according to the fluctuation period of the light intensity so that the segmentation frequency matches the light fluctuation frequency. Based on the incident angle information of the light parameters in the collected environment, the incident light projection angle of different regions on the surface of the aluminum-plastic composite material to be recycled is calculated. Differentiated segmentation similarity thresholds are set for regions with different projection angles. For regions where the reflection characteristics of the projection angle difference meet the preset reflection difference conditions, the segmentation similarity threshold matching the reflection difference conditions is adopted. A superpixel segmentation algorithm with a pre-set time sampling interval and a differentiated segmentation similarity threshold is used to segment the edge-enhanced visual acquisition data to obtain an initial set of superpixel units. A regional correlation analysis is performed on the initial superpixel unit set to calculate the texture similarity and color similarity of adjacent superpixel units. Adjacent superpixel units whose texture similarity and color similarity both exceed the set values ​​are merged, and isolated superpixel units with an area smaller than the minimum effective area are removed to form the superpixel unit cluster of the aluminum-plastic composite material to be recycled. For each superpixel unit in the superpixel unit cluster, the pixel grayscale value change sequence of the superpixel unit under the current illumination intensity and incident angle is collected, and the reflectivity change curve and refractive index change curve of the superpixel unit are calculated through the grayscale value change sequence. The reflectance and refractive index change curves of each superpixel unit are integrated according to the time dimension, and the spatial position coordinate information corresponding to the superpixel unit is supplemented to form the optical response data corresponding to each superpixel unit. A mapping index table is established between the superpixel unit cluster and the optical response data corresponding to each superpixel unit. The optical response data of any superpixel unit can be queried through the mapping index table. The optical response data reflects the reflection and refraction characteristics of the material under different lighting conditions. Based on the superpixel unit cluster and the optical response data corresponding to each superpixel unit, optical characteristic frequency domain analysis is performed to extract frequency domain feature parameters specific to aluminum-plastic composite materials. The material category determination model is constructed by combining the frequency domain feature parameters. The material category determination model is used to distinguish between aluminum component regions and plastic component regions in aluminum-plastic composite materials. The material category determination model is invoked to determine the category of the superpixel unit cluster and optical response data of the aluminum-plastic composite material to be recycled, and the composition partitioning result of the aluminum-plastic composite material to be recycled is obtained. The composition partitioning result includes the spatial location range of the aluminum component region and the spatial location range of the plastic component region. Based on the compositional zoning results of the aluminum-plastic composite material to be recycled, an aluminum-plastic separation and recycling instruction is generated. The aluminum-plastic separation and recycling instruction includes the sorting path planning of the aluminum component and the processing parameters of the plastic component, and the aluminum-plastic separation and recycling instruction is sent to the aluminum-plastic composite material recycling and sorting equipment.

2. The material recycling analysis method based on machine vision according to claim 1, characterized in that, The step involves calculating the incident light projection angle of different regions on the surface of the aluminum-plastic composite material to be recycled based on the incident light angle information in the illumination parameters of the collected environment, and setting differentiated segmentation similarity thresholds for regions with different projection angles, including: The three-dimensional morphological data of the aluminum-plastic composite material to be recycled is obtained. The three-dimensional morphological data includes the spatial coordinates and curvature information of each point on the material surface. The data is collected by a three-dimensional scanning device and converted into a standardized spatial coordinate matrix. Based on the incident angle information in the illumination parameters of the collected environment, the spatial direction vector of the incident light is determined, and the spatial direction vector of the incident light is used to perform vector operation with the spatial coordinate matrix in the three-dimensional morphological data to obtain the incident light projection angle at each point on the surface of the aluminum-plastic composite material to be recycled. The surface of the aluminum-plastic composite material to be recycled is divided into multiple angle intervals according to the incident light projection angle. Each angle interval contains a continuous projection angle range, and the proportion of material surface area contained in each angle interval is calculated. The optical reflection characteristics of aluminum-plastic composite materials are analyzed in different angle ranges. The reflectance difference between aluminum and plastic components at different projection angles is obtained through experimental data. The angle range where the reflectance difference meets the preset difference condition is matched with the corresponding segmentation similarity requirements. Based on the reflectance difference between aluminum and plastic components within each angular interval, a segmentation similarity threshold is set for that angular interval. When the reflectance difference meets the preset difference condition, the segmentation similarity threshold is set to the value that matches the difference condition, thus distinguishing aluminum and plastic components in areas where the reflectance difference meets the preset condition. The spatial range of each angle interval is mapped to the pixel coordinates of the visual acquisition data of the aluminum-plastic composite material to be recycled, so that each pixel can correspond to its own angle interval and the corresponding segmentation similarity threshold. During the superpixel segmentation process, for each region to be segmented, the angle interval to which the region to be segmented belongs is first determined, and the segmentation similarity threshold corresponding to the angle interval is called to make segmentation judgment, so that the segmentation process is adapted to the region projection angle. Record the usage of segmentation similarity thresholds and corresponding segmentation effects for different angle intervals. In subsequent analysis of similar material recycling, the segmentation similarity thresholds will be iteratively optimized based on historical segmentation effects.

3. The material recycling analysis method based on machine vision according to claim 1, characterized in that, The process involves performing frequency domain analysis of optical properties based on the superpixel unit cluster and the optical response data corresponding to each superpixel unit, extracting frequency domain feature parameters specific to the aluminum-plastic composite material, and constructing a material category determination model based on these frequency domain feature parameters, including: The optical response data corresponding to each superpixel unit in the superpixel unit cluster is preprocessed, and the preprocessed optical response data is converted into a time series signal. The time series signal has the acquisition time as the horizontal axis and reflectivity or refractive index as the vertical axis. Each superpixel unit corresponds to a set of reflectivity time series signals and a set of refractive index time series signals. Perform a Fourier transform on the time-series signal to convert the optical response change in the time domain into a spectral distribution in the frequency domain, thereby obtaining the optical characteristic spectrum of each superpixel unit. The optical characteristic spectrum contains amplitude and phase information of different frequency components. The spectral differences between aluminum and plastic components in aluminum-plastic composites were analyzed. The characteristic frequency ranges specific to aluminum and plastic components were determined through sample experiments. Each characteristic frequency range contains at least two reliably distinguishable frequency peaks. From the optical characteristic spectrum of each superpixel unit, spectral parameters within the characteristic frequency range are extracted. The spectral parameters include the amplitude of the characteristic frequency peak, the bandwidth of the characteristic frequency peak, and the total energy value within the characteristic frequency range. The above spectral parameters constitute the frequency domain characteristic parameters specific to aluminum-plastic composite materials. Collect superpixel unit clusters, optical response data and corresponding frequency domain feature parameters of different batches of aluminum-plastic composite materials to be recycled. Label the actual component category of each superpixel unit as aluminum or plastic, and construct a training dataset for the material category determination model. The training dataset is divided into a training sample set and a validation sample set. The training sample set is used for model parameter learning, and the validation sample set is used for model performance verification. The division ratio is determined based on the total number of samples. A network architecture for a material category determination model is constructed. The network architecture includes an input layer, a feature fusion layer, a classification layer, and an output layer. The input layer receives the frequency domain feature parameters of the superpixel unit. The feature fusion layer performs weighted fusion of different types of frequency domain feature parameters. The classification layer determines the component category through a fully connected neural network. The output layer outputs the component category prediction results of the superpixel unit. The training sample set is input into the network architecture, and the parameters of the material category determination model are adjusted by using the backpropagation algorithm. The weights of the material category determination model are optimized by calculating the loss value between the predicted result and the labeled category. The loss value is calculated using the cross-entropy loss function. The performance of the material category determination model during the training process is evaluated using the validation sample set. When the evaluation index remains stable and reaches the preset standard for multiple rounds of training, the model training is stopped, and the final material category determination model is obtained. The evaluation index includes the accuracy of component category determination, the recall rate of aluminum component identification, and the recall rate of plastic component identification.

4. The material recycling analysis method based on machine vision according to claim 3, characterized in that, The analysis of the spectral differences between the aluminum and plastic components in the aluminum-plastic composite material determined the characteristic frequency ranges specific to the aluminum and plastic components through sample experiments, including: We collected aluminum-plastic composite material samples with different material properties. The samples included aluminum-plastic composite materials with different aluminum content ratios and different types of plastics, namely polyethylene, polypropylene and polyvinyl chloride, so that the samples covered common aluminum-plastic recycling scenarios. Each aluminum-plastic composite material sample was physically separated to extract the aluminum and plastic components from the sample, resulting in pure aluminum and pure plastic samples. Under the same acquisition environment and lighting conditions, visual acquisition data and optical response data of pure aluminum samples and pure plastic samples were acquired respectively. During the acquisition process, the acquisition equipment parameters were kept consistent, namely focal length, exposure time, and photosensitivity. Following the same preprocessing procedure and Fourier transform parameters, the optical response data of pure aluminum samples and pure plastic samples were frequency domain transformed to obtain the spectrum diagrams of pure aluminum samples and pure plastic samples, respectively. By comparing the spectrograms of pure aluminum samples and pure plastic samples, we can find regions of significant difference in their spectral amplitude and frequency distribution, and record the frequency ranges corresponding to these regions of significant difference. These regions of significant difference are potential characteristic frequency ranges. For each potential characteristic frequency range, the mean and standard deviation of the spectral parameters of multiple groups of pure aluminum samples and pure plastic samples in that frequency range are calculated, and the degree of dispersion of the parameter distribution is analyzed. The frequency range in which the degree of dispersion meets the preset dispersion condition and the difference between the mean parameters of the two groups of samples meets the preset difference condition is the candidate characteristic frequency range. The statistical test method was used to verify the significance of the difference in spectral parameters between pure aluminum samples and pure plastic samples within the candidate characteristic frequency range. The significance level was set to a preset value. The candidate characteristic frequency range that passed the significance test was determined as the final characteristic frequency range. The final characteristic frequency range is divided into a characteristic frequency range specific to aluminum components and a characteristic frequency range specific to plastic components. The characteristic frequency range specific to aluminum components corresponds to a frequency range in which the spectral parameters of pure aluminum samples are significantly higher than those of pure plastic samples, while the characteristic frequency range specific to plastic components corresponds to a frequency range in which the spectral parameters of pure plastic samples are significantly higher than those of pure aluminum samples. The stability of the determined characteristic frequency range was verified. Under different light intensities and incident angles, the spectral data of pure aluminum samples and pure plastic samples were repeatedly collected to check whether the characteristic frequency range remained stable. If there were fluctuations, the frequency range boundary was adjusted. The aluminum-plastic composition was distinguished by this frequency range under different environments. Record the specific numerical ranges of the characteristic frequency ranges specific to aluminum components and plastic components, along with their corresponding spectral parameter characteristics and stability verification results, to form a characteristic frequency range document.

5. The material recycling analysis method based on machine vision according to claim 1, characterized in that, The step of generating an aluminum-plastic separation and recycling instruction based on the compositional zoning results of the aluminum-plastic composite material to be recycled includes: The composition zoning results of the aluminum-plastic composite material to be recycled are analyzed, the spatial location ranges of the aluminum component area and the plastic component area are extracted, and the spatial location ranges are converted into coordinate data consistent with the coordinate system of the recycling and sorting equipment, so that the recycling and sorting equipment can accurately locate the component area. The spatial distribution characteristics of aluminum component regions are analyzed, and the priority of aluminum components is divided according to the area size of the aluminum component regions. The spatial distribution characteristics include the number of aluminum component regions, the area size of each aluminum component region, and the relative positional relationship between each aluminum component region. The aluminum component regions whose area size meets the preset area conditions are set as the sorting priority matching the area conditions. Based on the coordinate data and priority of the aluminum composition region, combined with the movement range, movement speed and sorting accuracy parameters of the robotic arm of the recycling and sorting equipment, the sorting path of the aluminum composition is planned. Material property information of the plastic component region is extracted. The material property information is derived from the frequency domain feature parameters of the material category determination model. The material property information includes the melting temperature range, density parameter and chemical stability characteristics of the plastic component. The melting temperature range, density parameter and chemical stability characteristics of the plastic component are used to determine the processing technology of the plastic component. Based on the melting temperature range of the plastic components, determine the heat treatment temperature parameters of the plastic components, and control the heat treatment temperature within the melting temperature range so that the plastic components can be effectively melted without excessive degradation. Based on the density parameters of the plastic components, a corresponding separation process is adopted. The separation process is density sorting, which determines the density parameters of the sorting medium so that the plastic components are separated from other impurities through density differences. Based on the chemical stability characteristics of the plastic components, the process parameters for subsequent processing are determined. The subsequent processing processes are granulation and injection molding. The process parameters include processing pressure, processing time and cooling rate, so that the plastic components can be processed into qualified recycled raw materials. By integrating the sorting path planning data of aluminum components and the processing parameters of plastic components, and in accordance with the instruction format requirements of the recycling and sorting equipment, an aluminum-plastic separation and recycling instruction is generated. The aluminum-plastic separation and recycling instruction includes the instruction generation time, the identifier of the material to be processed, and the equipment operation sequence. Send the aluminum-plastic separation and recycling instruction to the aluminum-plastic composite material recycling and sorting equipment.

6. The material recycling analysis method based on machine vision according to claim 5, characterized in that, Based on the coordinate data and priority of the aluminum composition region, combined with the robotic arm's motion range, speed, and sorting accuracy parameters of the recycling and sorting equipment, the sorting path for the aluminum composition is planned, including: The motion parameters of the robotic arm of the recycling and sorting equipment are obtained. The motion parameters of the robotic arm include the maximum motion angle of each joint of the robotic arm, the upper limit of motion speed, the upper limit of motion acceleration, and the positioning accuracy error. The coordinate data of the aluminum composition region is converted into target coordinate points in the coordinate system of the robotic arm. Each aluminum composition region corresponds to at least one target coordinate point. The target coordinate point is selected as the geometric center of the aluminum composition region, so that the robotic arm can accurately grasp the aluminum composition region. Based on the priority of aluminum composition regions, the target coordinate points are sorted. The target coordinate points corresponding to aluminum composition regions that meet the preset priority conditions are placed at the front of the sorting sequence. At the same time, the sorting is adjusted based on the priority, taking into account the spatial distance between the target coordinate points, to reduce the round-trip movement distance of the robotic arm. Calculate the straight-line distance between two adjacent target coordinate points, and combine the upper limit of the robot arm's movement speed and the upper limit of its movement acceleration to estimate the theoretical movement time of the robot arm moving from one target coordinate point to the next target coordinate point. The theoretical movement time includes the total time of the acceleration phase, the constant speed phase, and the deceleration phase. Based on the positioning accuracy error of the robotic arm, a safe operating range is set around each target coordinate point. The size of the safe operating range enables the robotic arm to grasp the aluminum component area even in the presence of positioning error. The initial sorting path is planned and generated by connecting adjacent target coordinate points with straight lines according to the sorted target coordinate points. At the same time, it avoids obstacles within the range of movement of the robotic arm. The obstacles are the equipment frame and other materials. The location information of the obstacles is obtained from the environmental map of the recycling and sorting equipment. The initial sorting path is optimized by using a path smoothing algorithm to adjust the path curve and calculating the total length and total movement time of the optimized sorting path. The total movement time meets the industrial-grade detection throughput requirements, and the number of aluminum-plastic composite materials processed per unit time is not less than the preset throughput standard. If the total movement time exceeds the preset time range, the target coordinate point sorting is readjusted or the movement speed parameters are optimized. Record the coordinate sequence, motion speed curve, and motion time allocation of the final determined sorting path, and integrate them into sorting path planning data for aluminum composition.

7. The material recycling analysis method based on machine vision according to claim 1, characterized in that, The process of calling the material category determination model to classify the superpixel unit clusters and optical response data of the aluminum-plastic composite material to be recycled, and obtaining the composition partitioning results of the aluminum-plastic composite material to be recycled, includes: The optical response data corresponding to each superpixel unit in the superpixel unit cluster of the aluminum-plastic composite material to be recycled is converted according to the input format required by the material category determination model. The converted optical response data contains a standardized representation of the frequency domain feature parameters. The standardization process uses the standardized parameters used during the training of the material category determination model. The converted optical response data is input into the input layer of the material category determination model. The data is then dimensionally validated to ensure that the dimension of the input data matches the number of nodes in the input layer of the material category determination model. If the dimension does not match, a data format error message is returned, and the data conversion is performed again. The verified data is transmitted to the feature fusion layer. Different types of frequency domain feature parameters are weighted and summed according to preset weight coefficients. The different types of frequency domain feature parameters are the peak amplitude of the feature frequency, bandwidth, and total energy value, to obtain the fused feature vector. The weight coefficients are determined by the material category determination model training process and reflect the degree of contribution of different frequency domain features to the composition determination. The fused feature vector is transmitted to the classification layer, and the fused feature vector is subjected to nonlinear transformation through a fully connected neural network. Each neural network node processes the input data using a preset activation function, which is the ReLU function. The classification layer outputs the probability value of each superpixel unit belonging to the aluminum component and the probability value of belonging to the plastic component. The probability value ranges from 0 to 1, and the sum of the two probability values ​​is 1. When the probability value of a superpixel unit being made of aluminum is greater than the probability determination threshold, the superpixel unit is determined to be an aluminum component unit; when the probability value of a superpixel unit being made of plastic is greater than the probability determination threshold, the superpixel unit is determined to be a plastic component unit; if neither probability value is greater than the probability determination threshold, the superpixel unit is marked as a unit to be confirmed. For superpixel units marked as units to be confirmed, the detailed features in the optical response data of the superpixel unit are extracted again. The detailed features are the number of inflection points in the reflectance change curve and the frequency of refractive index fluctuation. These features are then added to the classification layer of the material category determination model, and the probability calculation and determination are performed again. Collect all superpixel units identified as aluminum components, analyze the spatial distribution of the superpixel units, merge adjacent aluminum components to form continuous aluminum component regions, and record the boundary coordinates and area of ​​each aluminum component region. Using the same method, adjacent plastic component units are merged to form continuous plastic component regions, and the boundary coordinates and area size of each plastic component region are recorded; The information of all aluminum and plastic component regions is integrated, including the number of regions, the boundary coordinates of each region, the area size, and the corresponding number of superpixel units, to form the component partitioning result of the aluminum-plastic composite material to be recycled; The composition partitioning results are converted into a visual image, and the boundary markers of the aluminum composition area and the plastic composition area are superimposed on the original visual acquisition data. Different colors are used to distinguish the different composition areas.

8. The material recycling analysis method based on machine vision according to claim 3, characterized in that, The step of inputting the training sample set into the network architecture, adjusting the model parameters using the backpropagation algorithm, and optimizing the model weights by calculating the loss value between the prediction result and the labeled category includes: The network parameters of the material category determination model are initialized. The network parameters include the weight coefficients of the feature fusion layer, the weight matrix of the fully connected neural network of the classification layer, and the bias vector. The initialization adopts a random normal distribution to keep the initial parameter values ​​within a reasonable range and avoid gradient explosion or gradient vanishing in the early stage of training of the material category determination model. The training sample set is divided into multiple training batches according to a preset batch size. Each training batch contains the same number of superpixel unit samples. The batch size is determined based on the memory capacity of the material category determination model training device. Starting from the first training batch, the sample data in the training batch is input into the input layer of the material category determination model. After processing by the feature fusion layer and the classification layer, the component category prediction probability of each sample is obtained. The component category prediction probability is the probability of aluminum component and the probability of plastic component. Based on the sample's label category (aluminum or plastic), construct a true label vector. If the sample is labeled as aluminum, the true label vector is 1, 0; if the sample is labeled as plastic, the true label vector is 0, 1. The cross-entropy loss function is used to calculate the loss value between the predicted probability and the true label vector of each sample. The cross-entropy loss function measures the difference between the two probability distributions. Calculate the average loss value for each training batch. The average loss value is the arithmetic mean of the loss values ​​of all samples in that training batch. The backpropagation algorithm is used to calculate the partial derivatives of the average loss value with respect to each network parameter, starting from the output layer of the material category determination model. Each network parameter is a weight coefficient, a weight matrix, and a bias vector. The partial derivatives are the gradient values, which reflect the direction and degree of influence of parameter changes on the loss value. Based on the calculated gradient values, the gradient descent optimization algorithm is used to adjust the parameters of each network. The parameter adjustment method is that the new parameter value is equal to the old parameter value minus the learning rate multiplied by the gradient value, and the learning rate is set to a preset initial value. After completing the parameter tuning for one training batch, proceed to the next training batch and repeat the steps of sample input, prediction calculation, loss calculation, gradient calculation and parameter tuning until all training batches have completed one training cycle. This process is considered one round of training. After each round of training, the validation loss value and component category determination accuracy of the substance category determination model are calculated using the validation sample set. The calculation method for the validation loss value and accuracy is the same as that for the training process. Compare the validation loss and accuracy of the current round with those of the previous round. If the validation loss decreases and the accuracy increases, continue to the next round of training. If the validation loss increases for several consecutive rounds or the accuracy remains unchanged for several consecutive rounds, the material category determination model training is considered to have reached convergence and training is stopped. During training, the training loss value, validation loss value, and accuracy of each training round are recorded, and the training curve is plotted. If overfitting occurs, where the training loss value decreases but the validation loss value increases, the dropout technique or L2 regularization method is used to adjust the material category determination model. After training stops, save the final material category determination model parameters and training process records. The final material category determination model parameters are the optimized weight coefficients, weight matrix, and bias vector.

9. A material recycling and analysis system based on machine vision, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the machine vision-based material recycling analysis method according to any one of claims 1 to 8 by executing the machine executable instructions.

Images