A method for screening and detecting microplastics in wastewater

Abstract

This invention discloses a method for screening and detecting microplastics in wastewater. The method includes: collecting wastewater samples; separating and collecting particulate samples from the wastewater samples, classifying the particulate samples into microplastics and non-microplastics; subjecting the collected particulate samples to deep oxidation treatment to generate oxidized functional groups on the surface of the microplastics; performing spectral detection on the deeply oxidized particulate samples using vibrational spectroscopy, selecting particulate samples whose spectra match the characteristic peaks of polymers, contain preset characteristic peaks of oxidized functional groups in their spectra, and conform to preset oxidation-bulk characteristic peak correlation rules; and labeling the selected particulate samples as microplastics. This invention improves the precision and accuracy of microplastic detection in wastewater.

Description

Technical Field

[0001] This invention belongs to the field of wastewater microplastic detection, and more specifically, relates to a method for screening and detecting microplastics in wastewater. Background Technology

[0002] Microplastic pollutants have been identified as a new class of environmental pollutants. Large amounts of microplastics are found in wastewater and sludge samples from urban wastewater treatment plants, and these microplastics ultimately end up in the environment, causing ecological harm. In practical applications, especially in the detection of complex environmental samples (such as sludge and sediments), current technologies still face many serious challenges.

[0003] (1) The problem of interference from complex matrices: In addition to microplastics, the particulate matter obtained by environmental sample screening also contains a large number of inorganic mineral particles, natural organic matter, metal oxides, etc. These components are often highly similar to microplastics in size and shape, resulting in a very high false judgment rate in the preliminary identification based on morphology optical microscopy, that is, the problem of false positives and false negatives is prominent. (2) Insufficient targeting of pretreatment steps: The existing sample pretreatment process (such as density separation and digestion) is mainly aimed at separating and purifying microplastics, but it lacks a chemical modification step that can actively enhance the specificity of microplastic detection. It is impossible to give microplastics a "specific label" that distinguishes them from background interferences through a controllable means before detection. (3) High requirements for detection instruments: The mainstream microplastic detection methods are optical methods and mass spectrometry. These two methods generally require professional laboratories and platforms, which are highly dependent on equipment and operators, and have low detection efficiency, making them difficult to promote.

[0004] Therefore, there is an urgent need in this field to develop a novel microplastic detection method that can effectively overcome interference from complex matrices, significantly improve identification specificity and reliability, and increase detection speed and efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for screening and detecting microplastics in wastewater, aiming to solve the problems of complex matrix interference, and the need to improve the specificity and reliability of identification in existing wastewater microplastic detection technologies.

[0006] Based on the above objectives, a first aspect of the present invention provides a method for screening and detecting microplastics in wastewater, the method comprising: S1, collecting wastewater samples, separating and collecting particulate samples from the wastewater samples, wherein the particulate samples are divided into microplastics and non-microplastics; S2, subjecting the collected particulate samples to deep oxidation treatment to generate oxidized functional groups on the surface of the microplastics; S3, performing spectral detection on the deeply oxidized particulate samples using vibrational spectroscopy, and selecting particulate samples whose spectra match the characteristic peaks of polymers, contain preset characteristic peaks of oxidized functional groups in the spectra, and conform to preset oxidation-bulk characteristic peak correlation rules; S4, labeling the selected particulate samples as microplastics.

[0007] Preferably, step S2 specifically includes: suspending the particulate sample in a hydrogen peroxide solution and irradiating the particulate sample in the solution with ultraviolet light; wherein the thickness of the suspension formed by the particulate sample suspended in the hydrogen peroxide solution is no higher than 2 cm, the peak value of the ultraviolet light is 200 nm to 280 nm, and the irradiation dose of the ultraviolet light is 3600 mJ / cm. 2 ~18000mJ / cm 2 The mass concentration of hydrogen peroxide solution is 1%~10%; or, the particulate sample is immersed in a potassium permanganate solution with a concentration of 0.01M~0.1M and reacted at a temperature of 20℃~60℃ for 10min~6h.

[0008] Preferably, the preset oxidation-bulk characteristic peak correlation rule includes: the oxidation degree index of the particulate sample is greater than a preset threshold; the oxidation degree index is A. 氧化峰 / A 参考峰 , where A 参考峰 A represents the stable characteristic peak intensity or area of ​​the bulk spectrum of the particulate sample before deep oxidation treatment. 氧化峰 The peak area or peak height of the oxidation characteristic peak in the differential spectrum is the difference between the spectrum after deep oxidation treatment and the spectrum before deep oxidation treatment.

[0009] Preferably, between S2 and S3, the process further includes: transferring the deeply oxidized particle sample onto filter paper and covering the filter paper with a positioning grid; scanning the filter paper with an optical microscope to record the position information of the particle sample in the positioning grid coordinate system and generating a particle sample distribution image; after S4, the process further includes: setting corresponding label information for microplastics and non-microplastics in the particle sample distribution image according to the labeling results in S4.

[0010] Preferably, the method further includes: acquiring multiple particle sample distribution images with label information as a dataset, and using the dataset to train a pre-set wastewater microplastic detection model, wherein the wastewater microplastic detection model is used to detect wastewater microplastic information in the target image.

[0011] Preferably, the wastewater microplastic detection model includes a feature extraction and enhancement module, a feature fusion and enhancement module, and a discrimination module connected in sequence. The feature extraction and enhancement module includes: a geometric morphology feature pathway for extracting and enhancing the geometric morphology features of the target object in the image; a biofilm feature pathway for extracting and enhancing the local texture features of the target object covered by biofilm in the image; and a detail texture feature pathway for extracting and enhancing the detail texture features representing the aging of the target object in the image. The feature fusion and enhancement module is used to fuse and enhance the geometric morphology features, the local texture features, and the detail texture features. The discrimination module is used to classify and discriminate the fused and enhanced features to determine whether the target object is wastewater microplastic.

[0012] Preferably, the geometric morphological features include features of fibrous structures, features of fragmented structures, and features of granular structures; the image processing procedure of the geometric morphological feature path includes: enhancing the edge orientation features of fibrous structures in the image using the histogram of directional gradients corresponding to the image; extracting the features of fibrous structures using depthwise separable convolution; adaptively extracting the features of fragmented structures in the image using a deformable convolutional network; using multi-scale circular convolutional kernels to simulate circular contours through weight distribution to enhance the response to near-circular closed boundaries, thereby extracting the features of granular structures in the image; enhancing the features of fragmented structures based on a spatial attention mechanism; and enhancing the features of granular structures by aggregating gradient features from the center to the edge based on radial max pooling.

[0013] Preferably, the extracted fibrous structure is characterized by:

[0014] ;

[0015] ;

[0016] in, The characteristics of the extracted fibrous structure, For dynamic convolution kernel weights, For local regions in the input feature map; The kernel size; As a learnable basic convolutional kernel, Here is the directional intensity function. Center pixel gradient direction, For the domain pixels gradient direction, To control the hyperparameters of orientation sensitivity, For the domain pixels The gradient magnitude.

[0017] Preferably, the image processing procedure of the biomembrane feature pathway includes: extracting biomembrane texture features at different levels using several convolutional layers with different dilation rates; fusing the biomembrane texture features at different levels and then performing feature enhancement to obtain biomembrane texture enhancement features; concatenating the features after adaptive average pooling of the biomembrane texture enhancement features with the features after adaptive max pooling of the biomembrane texture enhancement features to obtain concatenated features; and enhancing the features related to microplastic material in the concatenated features by learning channel weights to form the local texture features.

[0018] Preferably, the image processing procedure of the detailed texture feature pathway includes: performing cascaded max pooling on the image to enhance the spatial sensitivity of holes and cracks in the image; using Sobel edge detection to detect soft edge features of biofilms and hard edge features of aging cracks in the image after cascaded max pooling; using dilated convolution to detect dark area features inside holes and continuity features of long cracks in the image after cascaded max pooling; using pointwise convolution to detect local patch texture features in the image after cascaded max pooling; stitching together the features of each branch to generate a spatial attention map; and using a learnable biofilm gain factor to weightedly fuse the spatial attention map with the local texture features to obtain detailed texture features characterizing the aging of the target object.

[0019] Compared with the prior art, the advantages of the present invention include:

[0020] (1) A method for screening and detecting microplastics in wastewater is provided. Through deep oxidation treatment, the microplastics in wastewater are actively endowed with unique oxidative functional group characteristic peaks. The new signal is correlated with the polymer bulk peak using dual fingerprint verification technology. In principle, the influence of most inorganic and organic interferences in complex environmental matrices is eliminated. This solves the defects of traditional methods such as high false positive rate and inaccurate identification of aged microplastics, and improves detection speed and efficiency.

[0021] (2) A wastewater microplastic detection model is provided. In view of the characteristics of wastewater microplastic shape, surface biofilm attachment and surface aging, a new feature extraction enhancement module is designed for the model. The model uses three feature capture pathways, namely geometric morphology feature pathway, biofilm feature pathway and detail texture feature pathway, to capture the morphological features of the image. While preserving the integrity of the original feature information, the model’s ability to jointly perceive microplastic morphology, biofilm attachment state and aging traces is significantly enhanced, thereby improving the accuracy, precision and efficiency of subsequent wastewater microplastic detection. Attached Figure Description

[0022] Figure 1 A flowchart of a method for screening and detecting microplastics in wastewater provided in an embodiment of the present invention.

[0023] Figure 2This diagram illustrates the implementation process of the microplastic screening and detection method in wastewater provided in this embodiment of the invention.

[0024] Figure 3 The flowchart shows the Raman spectroscopy-specific detection and dual fingerprint verification algorithm provided in this embodiment of the invention.

[0025] Figure 4 The positioning grid provided in the embodiments of the present invention.

[0026] Figure 5 A flowchart illustrating the workflow of the geometric feature pathway provided in this embodiment of the invention.

[0027] Figure 6 This is a flowchart illustrating the process of a biomembrane characteristic pathway provided in an embodiment of the present invention.

[0028] Figure 7 A flowchart illustrating the workflow of the detail texture feature pathway provided in this embodiment of the invention.

[0029] Figure 8 A flowchart illustrating the workflow of the wastewater microplastic detection model provided in this embodiment of the invention.

[0030] Figure 9 Wastewater microplastics with different shapes or surface textures of pollutants are provided for embodiments of the present invention.

[0031] Figure 10 The shape attention heatmap provided in this embodiment of the invention introduces a feature extraction enhancement module.

[0032] Figure 11 The surface detail attention heatmap provided in this embodiment of the invention introduces a feature extraction enhancement module. Detailed Implementation

[0033] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0035] Furthermore, in the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "horizontal," "vertical," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] In the description of this specification, the references to terms such as "an embodiment," "a particular embodiment," or "the embodiment" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0037] Figure 1 This is a flowchart illustrating the method for screening and detecting microplastics in wastewater according to an embodiment of the present invention. (See attached document.) Figure 1 This method includes steps S1-S4, and its specific implementation process is as follows: Figure 2 As shown.

[0038] Step S1: Collect wastewater samples, separate and collect particulate samples from the wastewater samples, and classify particulate samples into microplastics and non-microplastics.

[0039] Step S2 involves performing deep oxidation on the collected particle samples to generate oxidized functional groups on the surface of the microplastics.

[0040] By selectively oxidizing the particulate sample, specific oxidized functional groups are introduced onto the surface of the microplastic, thereby providing a basis for spectral detection. In a preferred embodiment, step S2 specifically includes the following two implementation methods.

[0041] Deep oxidation treatment method 1—UV-Hydrogen Peroxide Deep Oxidation Method: The particulate sample is suspended in a hydrogen peroxide solution, and the sample is irradiated with ultraviolet light. The thickness of the suspension formed by the particulate sample in the hydrogen peroxide solution is no more than 2 cm, the peak wavelength of the UV light is 200 nm to 280 nm, and the UV irradiation dose is 3600 mJ / cm². 2 ~18000mJ / cm 2 The mass concentration of the hydrogen peroxide solution is 1% to 10%. The peak value of the ultraviolet light is, for example, 254 nm.

[0042] Specifically, the particulate sample is suspended in a 1%–10% hydrogen peroxide solution, maintaining a suspension thickness of <2 cm; the suspension is then placed under ultraviolet light (a low-pressure mercury lamp emitting ultraviolet light with a peak emission of 254 nm), with the ultraviolet intensity reaching the surface of the suspension set to 0.1 mW / cm². 2 ~1.2mW / cm 2 The irradiation time is 30-300 minutes. Ensure the irradiation dose of 254nm ultraviolet light is greater than 3600 mJ / cm². 2 The solution was stirred during the reaction. After the reaction was complete, particulate matter was obtained by filtration through glass fiber filter paper.

[0043] Deep oxidation treatment method 2—potassium permanganate oxidation method: Immerse the particulate sample in a potassium permanganate solution with a concentration of 0.01M~0.1M and react at a temperature of 20℃~60℃ for 10min~6h.

[0044] Specifically, the sample particles are immersed in a 0.01M~0.1M potassium permanganate solution and reacted at 20℃~60℃ for 10min~6h. After the reaction, sodium oxalate or sodium sulfite solution can be added to terminate the reaction and remove residual manganese dioxide. After the reaction, the particles are obtained by filtration through glass fiber filter paper. Step S2 can also be performed using other oxidation methods with similar effects or advanced oxidation methods.

[0045] Step S3: Vibrational spectroscopy is used to perform spectral detection on the particle samples after deep oxidation treatment, and particle samples that match the characteristic peaks of the polymer, have the preset characteristic peaks of the oxidation functional groups in the spectrum, and conform to the preset oxidation-bulk characteristic peak correlation rules are selected.

[0046] After oxidation treatment, vibrational spectroscopy was used to detect the characteristic peaks of oxidized functional groups, and the microplastics were identified by combining the characteristic peaks of the polymer bulk. Specifically, the following steps were taken: (1) Screening whether the spectrum of the target substance (i.e., the particle sample after deep oxidation treatment) matched the polymer characteristic peaks in the official spectral library, and initially guessing whether it was a microplastic polymer; (2) Screening whether there were preset characteristic peaks of oxidized functional groups in the spectrum. If so, the spectrum was further matched with the polymer bulk characteristic peak database to identify the candidate polymer type; (3) According to the pre-established oxidation-bulk characteristic peak association rule library, the correlation between the above oxidation peaks and the candidate polymer bulk peaks was verified. If the correlation was valid, the particle was determined to be the corresponding microplastic; if the correlation was invalid, it was excluded or marked as requiring manual review. The specific implementation process of steps S3 and S4 is as follows: Figure 3 As shown.

[0047] In a preferred embodiment, the preset oxidation-bulk characteristic peak correlation rule includes: the oxidation degree index of the particulate sample is greater than a preset threshold; the oxidation degree index is A. 氧化峰 / A 参考峰 , where A 参考峰 To determine the stable characteristic peak intensity or area of ​​the bulk spectrum of the particulate sample before deep oxidation treatment (e.g., for polypropylene, this could be selected at 1375 cm⁻¹). -1 The peaks are around 1600 cm⁻¹; for polystyrene, a peak around 1600 cm⁻¹ can be selected. -1 (benzene ring peaks on the left and right), A 氧化峰 Characteristic oxidation peaks in differential spectroscopy (e.g., C=O at 1710 cm⁻¹) -1 The peak area or peak height (left and right) is the difference between the spectrum after deep oxidation treatment and the spectrum before deep oxidation treatment. Satisfying A 氧化峰 / A 参考峰 If the value exceeds a preset threshold, the correlation is established. Differential spectroscopy will very clearly show only the changes resulting from the oxidation reaction: positive peaks represent newly formed functional groups (such as C=O), and negative peaks represent consumed functional groups (such as certain CH bonds). A 参考峰 As an internal standard, it corrects for differences caused by particle size and thickness. The higher the oxidation degree index, the higher the oxidation degree of the particle, which reflects its polymer type (some plastics are more prone to oxidation) and aging history.

[0048] The following describes the standard spectral database used in step S3.

[0049] Database 1—Bulk Characteristic Peak Library: Collects standard Fourier transform infrared and Raman spectra of target polymers (such as polyethylene, polypropylene, polystyrene, polyethylene terephthalate, polyvinyl chloride, nylon, etc.), and accurately marks the positions (wavenumber / cm) of their key bulk characteristic peaks. -1 For example, the key bulk characteristic peak of polyethylene is located at 1375 cm⁻¹. -1 The key bulk characteristic peak of polystyrene, located at 1600 cm⁻¹ (with methyl symmetry bending), is approximately 1600 cm⁻¹. -1 The key bulk characteristic peak of polyethylene terephthalate (PET) is located at 1710 cm⁻¹ (benzene ring skeletal vibration). -1 Left and right (C=O ester group).

[0050] Database 2—Oxidation Characteristic Peak Library: Standardized deep oxidation treatment (e.g., using uniform ultraviolet-hydrogen peroxide treatment under specific conditions) was performed on the same standard polymer particles. Then, their spectra were collected to establish an "Oxidized Polymer Spectral Library," accurately labeling the new characteristic peaks generated by oxidation, primarily those of carbonyl groups (C=O, at 1700 cm⁻¹). -1 ~1750cm -1), carboxyl group (-COOH, OH stretches at 3000cm) -1 ~3500cm -1 C=O at 1710cm -1 )wait.

[0051] Establish association rules: In the database, select several characteristic peak combinations and establish several association rules for each polymer. For example: Rule 1 - Polypropylene: Polypropylene bulk peak (1375 cm⁻¹) -1 1450cm -1 2800cm -1 ~3000cm -1 CH stretching) + carbonyl oxidation peak (1710 cm⁻¹) -1 Furthermore, its peak intensity conforms to a certain oxidation degree index value. Rule 2 - Polystyrene: Polystyrene bulk peak (1600 cm⁻¹) -1 1493cm -1 (and the out-of-plane bending vibration peaks of the benzene ring CH) + carbonyl oxidation peak (1710 cm⁻¹) -1 Furthermore, its peak intensity conforms to a certain oxidation degree index value.

[0052] The specific matching process based on the above database is as follows.

[0053] Bulk peak matching: Analyze the spectrum and match it with Database 1 (bulk characteristic peak library) to identify the most likely polymer types (matching rate >70%).

[0054] Oxidation peak screening: Search the entire spectrum for obvious oxidation characteristic peaks (such as at 1700 cm⁻¹). -1 ~1750cm -1 (Whether a new peak appears within the range).

[0055] Correlation verification: Check whether the ratio of the detected oxidation peak intensity to the matched polymer bulk peak intensity is "reasonable". For example: whether the shape and position of the detected oxidation peak are consistent with the oxidation peak shape of this polymer in Database 2. If there is a very strong 1710 cm⁻¹ peak in the spectrum... -1 The peak was detected, but no bulk peak of any polymer was found. The algorithm should classify it as a "suspicious interfering substance" (possibly some natural organic matter) rather than microplastics.

[0056] Step S4: The selected particle samples are labeled as microplastics.

[0057] The output results are as follows: "Microplastic type: polypropylene; Confidence level: high (both bulk and oxidation peaks were detected, and the oxidation degree index value is within a reasonable range)"; or "Unknown particles: suspected bulk peaks were detected, but no oxidation peaks were detected. Manual verification is recommended"; or "Unknown particles: oxidation peaks were detected, but no matching bulk peaks were detected. Manual verification is recommended".

[0058] In a preferred embodiment, between steps S2 and S3, the method further includes: transferring the deeply oxidized particle sample onto filter paper and covering the filter paper with a positioning grid; scanning the filter paper using an optical microscope to record the position information of the particle sample in the positioning grid coordinate system, thereby generating a particle sample distribution image. After step S4, the method further includes: setting corresponding label information for microplastics and non-microplastics in the particle sample distribution image based on the labeling results in step S4.

[0059] In a preferred embodiment, the method further includes: acquiring multiple particle sample distribution images with label information as a dataset, and using the dataset to train a pre-set wastewater microplastic detection model, wherein the wastewater microplastic detection model is used to detect wastewater microplastic information in the target image.

[0060] Existing wastewater microplastic detection models have limited ability to identify microplastics from specific sources. This is mainly due to two factors: firstly, the sources of original microplastic image data for their training sets are limited, and there is currently no image database related to microplastics; secondly, the resulting neural network deep learning frameworks have not been deeply developed and specifically optimized for microplastics with specific morphological features, resulting in poor generalization ability and affecting the accuracy of microplastic detection. In reality, microplastic image features vary greatly. For example, the morphology of marine microplastics differs significantly from that of terrestrial river microplastics, and microplastics in water bodies differ from those in soil. Key factors such as the source, material, aging process, and degree of aging of microplastics all contribute to their morphological differences. Currently, there are no specific detection methods established for microplastics in wastewater and sludge from wastewater treatment plants. Existing conventional microplastic detection methods are insufficient in accuracy, precision, and efficiency when directly applied to wastewater microplastic detection.

[0061] Microplastics exist in various forms in wastewater and sludge, but their shapes mainly include fibrous, granular, and fragmented forms, which are closely related to their sources. Fibrous (filamentous) microplastics mainly originate from microplastics shed during clothing washing; granular (especially spherical) microplastics mainly come from microplastics released from everyday personal care products; fragmented microplastics often exhibit irregular shapes, and their sources are more complex, potentially originating from various industrial production processes and daily life. Microplastics in oceans and rivers are primarily in the form of fragments and films of various irregular shapes, which are also closely related to their sources, with significant differences in shape and detailed characteristics among microplastics from different sources.

[0062] In addition to the above, during wastewater treatment, microplastics may undergo specific aging processes, such as aging caused by disinfection and biodegradation, or they may have various specific microorganisms and their derived organic matter attached to their surfaces, such as extracellular polymers. Wastewater treatment processes mainly include biological, chemical, and physical methods, all of which affect the surface morphology of microplastics. Biological methods mainly include activated sludge and biofilm processes; after biological treatment, microplastics may have microbial films or extracellular substances of microorganisms attached to their surfaces. Chemical methods mainly include coagulation and chlorination; coagulants added during coagulation can also adhere to the microplastic surface, while chlorination has a certain corrosive effect on microplastics. Physical methods mainly involve sedimentation and ultraviolet (UV) disinfection. Sedimentation has a relatively small impact on microplastics, while UV disinfection may cause aging of the microplastic surface, thus degrading it. This results in a significant difference in morphology between microplastics from wastewater treatment plants and those from other sources. Therefore, the same models and methods cannot be used to identify microplastics from different sources.

[0063] To address the aforementioned shortcomings, this invention designs an improved wastewater microplastic detection model based on technologies such as machine learning, reinforcement learning, and deep neural networks. The wastewater microplastic detection model comprises a feature extraction enhancement module, a feature fusion enhancement module, and a discrimination module connected in sequence.

[0064] The feature extraction and enhancement module includes: a geometric morphology feature pathway, used to extract and enhance the geometric morphology features of the target object in the image; a biofilm feature pathway, used to extract and enhance the local texture features of the target object covered by biofilm in the image; and a detail texture feature pathway, used to extract and enhance the detail texture features in the image that characterize the aging of the target object.

[0065] The feature fusion and enhancement module is used to fuse and enhance geometric features, local texture features, and detailed texture features; the discrimination module is used to classify and discriminate the fused and enhanced features to determine whether the target object is wastewater microplastics.

[0066] This invention proposes a novel Microplastic-Specific Morphology Attention Enhanced Module (MSMAEM), a feature extraction enhancement module, which uses targeted enhancement of three feature capture pathways to capture the morphological features of patterns, targeting the characteristics of microplastics in wastewater, including their shape, surface bioattachment, and surface aging.

[0067] The first feature pathway is the geometric morphology feature pathway. Microplastics exhibit diverse morphologies in wastewater and sludge, but their geometric morphologies mainly include fibrous, granular, and fragmented forms. The purpose of this pathway is to enhance the model's attention to these geometric morphologies.

[0068] The second feature pathway is the biofilm feature pathway. Biofilm coverage changes local reflectivity but does not significantly alter the overall color mean, necessitating enhancement of local texture response. This pathway aims to explicitly capture the local texture changes caused by biofilm, thereby improving the model's ability to recognize biofilm layers on microplastic surfaces.

[0069] The third feature path is the texture detail feature path. Aging and eroded microplastic surfaces exhibit aging traces, manifested as high-frequency spatial discontinuities. The purpose of this path is to improve the model's sensitivity to micro-fractures / holes and enhance its response to discontinuous edges such as holes and cracks.

[0070] In a preferred embodiment, the geometric morphological features include features of fibrous structures, fragmented structures, and granular structures. The image processing procedure for the geometric morphological feature pathway includes: enhancing the edge orientation features of fibrous structures in the image using the histogram of oriented gradients corresponding to the image; extracting the features of fibrous structures using depthwise separable convolution; adaptively extracting the features of fragmented structures in the image using a deformable convolutional network; simulating circular contours through weight distribution using multi-scale circular convolutional kernels to enhance the response to near-circular closed boundaries, thereby extracting the features of granular structures in the image; enhancing the features of fragmented structures based on a spatial attention mechanism; and aggregating gradient features from the center to the edge based on radial max pooling to enhance the features of granular structures. In this embodiment, multi-scale circular convolutional kernels are used to simulate circular contours through weight distribution, effectively enhancing the response to near-circular closed boundaries; simultaneously, radial max pooling is used to aggregate gradient features from the center to the edge to strengthen the isoaxiality and closure representation of the particles.

[0071] In a preferred embodiment, the fibrous structure extracted by the geometric morphology feature pathway has the following characteristics:

[0072] ;

[0073] in, The characteristics of the extracted fibrous structure, For dynamic convolution kernel weights, For local regions in the input feature map; The kernel size; As a learnable basic convolutional kernel, Here is the directional intensity function. Center pixel gradient direction, For the domain pixels gradient direction, To control the hyperparameters of orientation sensitivity, For the domain pixels The gradient magnitude. In this embodiment, the weights of the dynamic weighted convolution kernel are determined by the directional relationship between the center pixel and its neighboring pixels. It is a preset hyperparameter, and its valid value range is usually within 1000. between.

[0074] Preferably, in this invention, morphological prior knowledge is injected into the learning of the offset of deformable convolution, utilizing the deformable convolutional network to adaptively extract features of fragmented structures in the image. Offset modulation function. for:

[0075] ;

[0076] in, This is the original learnable offset; It is a learnable scalar used to control the strength of priors; Let be a rotation matrix, and let its rotation angle be... It is predicted from a global context vector (e.g., features obtained through pooling layers), which encodes the orientation of dominant edges in the image (e.g., horizontally indicating fibers, multiple orientations indicating fragments). It is a fixed unit direction vector, for example, initially radial. It is obtained through end-to-end learning using the standard backpropagation algorithm, and its value range is: However, in practice, due to constraints in weight initialization and regularization, Absolute values ​​are usually limited to a reasonable range, for example Within a distance of one pixel, to ensure sampling stability. As part of the model parameters, they are automatically optimized during training using gradient descent. The initial value is usually set to 0.2, 0.5, etc., to balance the task loss and corner prior loss in the early stage of training.

[0077] Preferably, in the feature extraction process of granular structures, the ring weight function of the multi-scale ring convolution is:

[0078] ;

[0079] in, Midpoint of the convolution kernel The final weight, For point Distance to the center of the convolution kernel; The target ring radius is a learnable parameter used to match particles of a specific size. The loop width control parameter controls the concentration of the weight distribution; it can be learned or fixed. The initial value is set according to the scale of the convolution kernel. For example, for a set of 3 kernels, the corresponding target ring radius can be initialized to 1.5, 3.0 and 4.5 respectively. The network learns the most discriminative particle contour radius in different datasets through training. It is a learnable parameter, with an initial value, for example, set to 0.5.

[0080] Preferably, in the feature enhancement process for granular structures, the radial max pooling operation, sector feature extraction, isoaxiality / roundness scoring, and final output are used. They are respectively:

[0081] ;

[0082] in, The maximum eigenvalue within the nth sector region; This represents the nth sector region centered on the candidate point; This indicates calculating the maximum value of all sectors. standard deviation To compress the standard deviation to Functions within an interval; For isometric scoring, the smaller the standard deviation, the higher the score; the closer to 1, the greater the probability that the point is the center of a perfect circle. This indicates a splicing operation. Indicates will Concatenate them sequentially to form an n-dimensional feature vector; It is a small neural network used to fuse features from all sectors.

[0083] The geometric morphology feature pathway specifically models the diverse morphologies of microplastics in wastewater and sludge. Wastewater treatment plant-derived microplastics primarily exhibit three main geometric morphologies: fibrous (slender linear structures), granular (quasi-spherical or spherical), and fragmented (irregular polygonal). For details, please refer to [link to relevant documentation]. Figure 5The geometric morphology feature pathway first enhances the edge orientation features of the fibrous structure through the histogram of oriented gradients (HARQ) in the convolutional layer, and then uses depthwise separable convolution to extract the fiber stump features (such as aspect ratio) of the microplastics. Next, a deformable convolutional network adaptively matches the geometric topology of fragmented, irregular contours, dynamically learning the convolutional kernel offset to allow the receptive field to flexibly adapt to boundary changes of different shapes. Subsequently, in the feature enhancement stage, a channel-spatial dual attention mechanism works collaboratively. The channel attention layer enhances the global saliency of granular closed contours, while the spatial attention layer focuses on the corner distribution features of fragmented polygons, thus forming a robust recognition capability for microplastic morphological variations.

[0084] In a preferred embodiment, the image processing procedure for the biomembrane feature pathway includes: extracting biomembrane texture features at different levels using several convolutional layers with different dilation rates; fusing the biomembrane texture features at different levels and then performing feature enhancement to obtain biomembrane texture enhancement features; concatenating the features obtained by adaptive average pooling of the biomembrane texture enhancement features with the features obtained by adaptive max pooling of the biomembrane texture enhancement features to obtain concatenated features; and enhancing the features related to microplastic material in the concatenated features by learning channel weights to form local texture features.

[0085] Multi-scale texture response fusion function is used in biomembrane feature pathways, let F d1 F d2 F d3 Given the output features of three dilated convolution branches with dilation rates of 1, 2, and 3 respectively, the scale-specific texture responsivity... for:

[0086] ;

[0087] in, It is feature map F d (i.e. F) d1 F d2 F d3 In The standard deviation is calculated within a small neighborhood of the center. For feature map F d height, For feature map F d width, For feature map F d The number of channels.

[0088] Adaptive fusion weights and final fusion output They are respectively:

[0089] ;

[0090] in, This is a temperature hyperparameter used to soften the weight distribution; For the first Texture responsivity of feature maps at each scale; The index variable represents all scales. ; For the first The feature map output by the dilated convolution branches at each scale. The optimal value is determined by performing a grid search on the validation set, and its effective range is typically within a certain range. between.

[0091] This study delves into the complex texture changes resulting from biofilm adhesion on microplastic surfaces, analyzing characteristic pathways of biofilms. Biofilms typically contain microscopic features such as mucus-like structures, bacterial dot colonies, and air bubbles. For details, please refer to [link to relevant documentation]. Figure 6 The biofilm feature pathway employs a multi-scale dilated convolutional architecture for hierarchical texture extraction. Parallel processing with convolutional layers of varying dilation rates—for example, using convolutional layers with dilation rates of 1, 2, and 3—captures multi-scale features ranging from local texture details to macroscopic community distribution. Subsequently, in the channel attention pathway, adaptive average pooling and max pooling feature inputs are fused, and channel weights are learned through a fully connected layer with a bottleneck structure. This approach focuses on enhancing the unique material properties of microplastics (such as reflectivity, transparency variations, and aging-induced differences in translucency) while effectively suppressing redundant channel responses associated with background inorganic particles, significantly improving material discriminability.

[0092] In a preferred embodiment, the image processing procedure of the detail texture feature pathway includes: performing cascaded max pooling on the image to enhance the spatial sensitivity of holes and cracks in the image; using a multi-branch edge-aware architecture, extracting multiple branch features representing the aging of the target object in the image after cascaded max pooling, and splicing the branch features to generate a spatial attention map; and using a learnable biomembrane gain factor to weightedly fuse the spatial attention map with the local texture features to obtain the detail texture features representing the aging of the target object.

[0093] In a preferred embodiment, the detailed texture feature pathway utilizes a multi-branch edge-aware architecture to extract multiple branch features representing the aging of the target object within the cascaded max-pooling image. Specifically, this includes: using Sobel edge detection to detect soft edge features of biofilms and hard edge features of aging cracks within the cascaded max-pooling image; using dilated convolution to detect dark area features inside holes and continuity features of long cracks within the cascaded max-pooling image; and using pointwise convolution to detect local patch texture features within the cascaded max-pooling image. In this embodiment, pointwise convolution (1×1 convolution) is used to perform cross-channel integration and dimensionality reduction of preceding features, specifically for analyzing and learning the inter-channel correlation characteristics of local patch textures caused by aging, generating a high-dimensional patch saliency feature map.

[0094] The cascaded max pooling used in the detail texture feature path leverages the difference operation to highlight discontinuities. Let the input feature map be X; the first stage pooling... Secondary pooling (stride=1, padding=1); Differential feature extraction:

[0095] ;

[0096] in, This is an upsampling function used to enlarge the size of the input feature map to a specified target size; The size of the feature map after the first stage of max pooling; for Shape attributes; This indicates a smaller feature map. Amplified to the same size using methods such as bilinear interpolation They have the exact same height and width so that precise element-by-element subtraction can be performed later.

[0097] The biofilm gain factor α should not be a fixed value, but should be adaptively adjusted according to the salience of the biofilm in the input image. Biofilm salience measurement. and adaptive gain factor They are respectively:

[0098] ;

[0099] in, It is the output of the texture branch. It is an indicator function. It is a threshold. The biofilm saliency measure formula calculates the proportion of pixels judged as salient texture. The base gain that can be learned; and It is a hyperparameter used to control the amplitude and sensitivity of the adjustment. Automatic optimization during training is achieved through backpropagation, typically by initializing the algorithm to fall within the expected range. Within a reasonable range. Used to control the maximum adjustment range of the gain factor, with a value range of [value missing]. . Used to control the sensitivity of the function to changes in texture complexity, with a value range of [value range missing]. .

[0100] The detailed texture feature pathway is specifically designed to detect microplastic aging traces and boundary features. For details, please refer to [link to documentation / reference]. Figure 7 The detailed texture feature pathway first enhances the spatial sensitivity to micropores and cracks through a cascaded max pooling strategy (e.g., 2×2 followed by 3×3 cascades), significantly improving the detection capability of aging traces. Then, a three-branch edge-aware architecture is employed: the first branch uses a fixed Sobel operator to strengthen the gradient difference response between boundaries (soft edges) and aging cracks (hard edges); the second branch expands the receptive field through convolutions with a hole ratio of 2, specifically capturing the continuity features of dark areas inside pores and long cracks; the third branch uses 1×1 convolutions to analyze local patchy textures caused by aging. The outputs of the three branches are spliced ​​and fused to generate a spatial attention map, and an innovative learnable biomembrane gain factor α is introduced to dynamically weight and fuse the biomembrane mask (i.e., the local texture features output by the biomembrane feature pathway) with the spatial attention map, generating an enhanced attention map for biomembrane perception, significantly improving the localization accuracy of key discrimination regions such as biomembrane-covered areas, pore structures, and crack pathways. The biofilm mask is a single-channel probability map with the same spatial resolution as the input feature map. Through its internal multi-scale dilated convolution and channel attention mechanism, the mask autonomously learns and highlights areas in the image that may be covered by biofilm. The higher the value, the higher the confidence that the area has biofilm features. This design achieves adaptive synergistic enhancement of biofilm features with texture and shape features.

[0101] In a preferred embodiment, the fusion enhancement process of the feature fusion enhancement module includes: concatenating geometric features, local texture features, and detailed texture features along the channel dimension, and then performing feature-level fusion through a convolutional layer; adding the fused features to the original input features in the image through residual connections to obtain the fused and enhanced features.

[0102] In this embodiment, the feature fusion and enhancement mechanism adopts a multi-level attention synergy architecture. Preferably, the outputs of the three feature pathways are integrated in the following way: the output of the geometric morphology feature pathway is upsampled to the original spatial resolution through bilinear interpolation; the channel attention weights of the biofilm feature pathway output are directly applied to the input features; and the biofilm perception enhancement attention map output by the detail texture feature pathway provides feature calibration in the spatial dimension. After the three features are concatenated in the channel dimension, they are fused at the feature level through convolutional layers, and finally, the fused features are added to the original input features through residual connections. This design significantly enhances the model's ability to jointly perceive microplastic morphology, biofilm attachment status, and aging traces while preserving the integrity of the original feature information, providing a highly discriminative feature representation for the subsequent classifier.

[0103] In this embodiment, the feature fusion enhancement module performs three-path attention collaborative fusion of geometric morphology features, local texture features, and detailed texture features. The following settings are made: the spatial attention map output by the geometric morphology feature path is... The channel attention vector output by the biomembrane characteristic pathway is The spatial attention map output by the detail texture feature path is as follows: The original input features are This approach employs a channel-space cross-attention mechanism for fusion, rather than simple point-by-point multiplication or concatenation. (Collaborative Attention Map Generation) for:

[0104] ;

[0105] in, A lightweight convolutional or fully connected layer for channel attention. Map to a dimension that matches the spatial graph or perform feature transformation; This represents an interactive operation; for example, it could be a Hadamard product, or it could be designed as... That is, the spatial graph is weighted using the transformed channel weights; A fusion function (e.g., sigmoid) is used to ensure that the values ​​in the final attention map are between 0 and 1. Final enhanced output. .

[0106] The fusion formula in this embodiment has the following advantages.

[0107] (1) Deep interaction. Instead of simply adjusting the weights of channels and the positions of spatial attention maps, channel attention information is incorporated into the fine-tuning of the two spatial attention maps. For example, if a channel is judged to be important to "biomembrane", the biomembrane region in the spatial attention map corresponding to that channel will be further enhanced.

[0108] (2) Weighted summation. The two spatial diagrams representing “form” and “detail” are combined in a weighted summation manner, and the interaction with channel information is used to achieve the purpose of synergistic enhancement.

[0109] (3) Residual form. Using By performing residual learning, the network can retain its original features even when the attention map is not significant, resulting in more stable training.

[0110] In a preferred embodiment, the discrimination module uses a global hybrid pooling strategy to aggregate spatial information from the features output by the feature fusion enhancement module. After dropout regularization, the classification result is output through a fully connected layer. The global hybrid pooling strategy includes, for example, a weighted combination of average pooling and max pooling; the classification result includes, for example, bounding boxes for wastewater microplastics and non-microplastics, category labels, and confidence scores.

[0111] In this embodiment, sample data needs to be collected before training the wastewater microplastic detection model. Another specific process for collecting sample data includes, for example, the following steps 1-6.

[0112] Step 1: Collect wastewater treatment plant samples for microplastic extraction. Specifically, the collected wastewater treatment plant samples, including water samples and sludge samples, are processed using specialized separation methods for microplastics. Separation methods include, but are not limited to, digestion, filtration, density separation, and sedimentation. After separation, the particulate matter in the samples is separated and collected on glass fiber filter paper.

[0113] Step 2: Using a Raman spectrometer equipped with a microscope, the microplastic samples on the glass fiber filter paper are marked and their materials are identified, and corresponding microscopic images are obtained.

[0114] First, adopt such as Figure 4 The homemade grid shown has a mesh size of 0.5 cm (greater than or equal to the maximum size limit of microplastics) and a line width of 0.1 cm. The grid can be made using 3D printing or photolithography. The grid material is not limited and can be selected according to experimental needs. A coordinate system containing alphanumeric symbols is formed on the grid using photolithography or 3D printing. Its size is just enough to cover the filter paper. The prepared grid is then placed on top of the filter paper and fixed in place.

[0115] Gridded filter paper is scanned under an optical microscope to record the precise position of target particles in the grid coordinate system and create a sample distribution map. The grid is then transferred to the sample stage of a micro Raman spectrometer. Based on the recorded position information, the spectrometer navigates to the target particle and acquires spectra, while simultaneously recording the coordinates of the spectrometer sample stage. Alternatively, the gridded filter paper can be sputter-coated with gold and then transferred to a scanning electron microscope for observation, obtaining high-resolution morphological images and energy-dispersive X-ray elemental spectra. By combining the Raman spectra, scanning electron microscopy morphology, and elemental data of the same particle, correlation analysis is performed to complete the final identification and characterization of the microplastics.

[0116] The above methods enable cross-platform, traceable, and precise analysis of individual microplastic particles, organically combining chemical identification with physical morphology and elemental analysis for mutual verification. This greatly eliminates the uncertainties that may arise from single-type detection technologies and significantly improves the depth and authority of microplastic detection and analysis.

[0117] Step 3: Preprocessing of microplastic contaminant images. The acquired images undergo preprocessing and feature extraction to specifically optimize and enhance their features. First, the acquired images are segmented using appropriate methods to obtain images with the target object as the main subject. The image resolution is 512×512 pixels, with a resolution of 80~120 dots per inch. Then, the image features are optimized. The target object's outline is drawn in the image processing software, and its resolution and contrast are adjusted to highlight the texture and outline features of the target object image. Finally, based on the Raman spectroscopy analysis results, the images are labeled. For example, microplastic target object files are labeled M1, and non-microplastic target object files are labeled M2.

[0118] Step 4, Data Loading and Labeling. Specifically, image data for the training and test sets are loaded from the specified paths. The data uses a binary classification labeling system (M1 / M2) to distinguish between microplastics (M1) and non-microplastics (M2), and numerical labels (0 or 1) are generated by mapping filenames.

[0119] Step 5, Data Feature Enhancement and Standardization. A composite data feature enhancement process is applied, including random rotation (±10° range), random horizontal flipping, and color jittering (e.g., adjusting brightness, contrast, saturation, and hue). The graphic data is converted into tensors and normalized to the [0,1] interval, then standardized to the [-1,1] range, forcing single-channel images to be converted to a red-green-blue (RGB) three-channel format.

[0120] Step 6, Dataset Encapsulation and Loading. Implement a custom dataset class, integrating image loading, label parsing, and preprocessing. Generate batch data (batch size = 8) using the DataLoader, supporting random shuffling of the training set.

[0121] The collected sample data is divided into training and validation sets. A wastewater microplastic detection model is trained with the goal of minimizing the model's loss function. Preferably, a cross-entropy loss function and adaptive optimization algorithm are used during the training phase, while the validation phase employs multi-dimensional evaluation metrics (accuracy, confusion matrix, and rate of change curve) to ensure the model's robustness in complex wastewater treatment scenarios, providing a reliable technical solution for accurate identification of wastewater microplastics.

[0122] For example, the loss function can be implemented using cross-entropy loss (for binary classification tasks), and the optimizer can be the Adam adaptive moment estimator. The learning rate is set to 2e-5, and the momentum parameter β is dynamically scheduled (0.5, 0.999). Learning rate decay is triggered based on the validation loss, with a decay factor of 0.1 and a tolerance epoch of 5. An early stopping mechanism (tolerance epochs = 10, saving the best model) can be used to shorten unnecessary training time. Iterative training (e.g., 100 epochs, which can be increased or decreased according to actual conditions) trains the discriminator using only labeled data, calculates the classification loss, and updates the weights through backpropagation. Monitoring metrics include recording the discriminator loss and classification accuracy for each batch, and saving model checkpoints every 5 epochs. In the validation phase, model performance is evaluated on an independent test set, generating a loss / accuracy curve. The overall process of sample data collection, wastewater microplastic detection model sample data processing, model training, and model validation is as follows: Figure 8 As shown.

[0123] The wastewater microplastic detection model provided by this invention adopts an innovative three-path feature collaborative architecture, deeply integrating discriminative information from geometric morphology features, local texture features (i.e., biofilm characteristics), and detailed texture features (i.e., aging traces). The discriminator uses a ResNet50 backbone network as the basic feature extractor, integrating a feature extraction enhancement module on its last layer's 2048-channel feature map. The feature extraction enhancement module achieves feature optimization through three parallel processing pathways: the geometric morphology feature pathway uses a deformable convolutional network and an oriented gradient feature enhancement layer to adaptively capture the geometric topological features of fibrous, granular, and fragmented microplastics; the biofilm feature pathway uses a multi-scale dilated convolutional architecture to parse the hierarchical texture of surface attachments, combined with a dual-pooling channel attention mechanism to enhance the discriminative power of material properties; and the detailed texture pathway uses a cascaded pooling strategy and a three-branch edge-aware architecture to accurately locate aging cracks and pore structures, and introduces a learnable biofilm gain factor to dynamically adjust the response intensity of the biofilm region.

[0124] The effectiveness of the wastewater microplastic detection model provided by this invention will be further explained in detail below, using the basic Squeeze-and-Excitation (SE) attention module.

[0125] The feature extraction enhancement module in the wastewater microplastic detection model constructed in this invention is replaced with a basic SE attention module. This basic SE attention module only contains multiple convolutional layers and does not have the feature extraction enhancement module of this invention. Furthermore, according to the model framework diagram, the discriminator module has a significant impact on the overall model's recognition accuracy. Therefore, several built-in networks are selected, including ResNet50, Residual Network 101 (ResNet101), and EfficiencyNet. The model is trained on the same batch of microplastic image data training set (approximately 3000 images) and validation set (approximately 700 images), and the model is generalized to the test set (200 microplastic images). During the above model training process, the Dropout layer inactivation rate is selected in the range of 0.1-0.5; the initial learning rate is selected in the range of 0.00004-0.0004; and the number of consecutive rounds of the early stopping mechanism is 5-20 rounds. Through repeated training and experimentation, the optimal parameters were finally determined as follows: Dropout layer inactivation rate of 0.2; initial learning rate of 0.0002, with dynamic learning rate adjustment based on training loss; and an early stopping mechanism for 10 consecutive rounds, stopping if the loss does not improve after 10 consecutive rounds. Examples of images used for training are provided below. Figure 9 As shown.

[0126] The training results are as follows: With the ResNet50 preset network, the microplastic recognition accuracy is 79%; with the ResNet101 preset network, the accuracy is 76%; and with the EfficiencyNet preset network, the accuracy is 70%. Based on these results, it can be seen that the ResNet series networks have significantly higher accuracy in recognizing microplastics than the EfficiencyNet series networks. Furthermore, under the same conditions and parameters, ResNet50 performs slightly better than ResNet101. Since ResNet101 has deeper network layers than ResNet50, its training consumes more computational resources; therefore, the ResNet50 network is preferred. In conclusion, it can be seen that the basic model based on the SE attention module has an accuracy of less than 80% for recognizing microplastic samples, indicating significant room for improvement.

[0127] The wastewater microplastic detection model proposed in this invention is trained using the same graphical data as the basic model based on the fundamental SE attention module described above. The shape attention heatmap of the model in this invention is shown below. Figure 10 As shown, the surface detail attention heatmap is as follows: Figure 11 As shown, the trained model was used for image recognition, achieving a 93% accuracy rate in recognizing microplastic images. This represents a significant improvement in accuracy compared to the model trained using the basic SE attention module.

[0128] Furthermore, the target image is input into the trained wastewater microplastic detection model to obtain information about wastewater microplastics in the target image.

[0129] The wastewater microplastic detection method uses a trained wastewater microplastic detection module to quickly screen and distinguish between microplastic pollutants and non-microplastic impurities in environmental samples. The specific detection process includes, for example: First, loading the trained wastewater microplastic detection model and maintaining the same image preprocessing workflow as during training (resize adjustment → tensor transformation → standardization); then performing the inference process. For single-image analysis, this includes preprocessing the input image (adding batch dimensions), outputting classification probabilities through model forward propagation, and normalizing the predicted category and confidence score using the Softmax function. For batch processing, multiple images are loaded from a specified directory, and classification inference is performed in parallel, outputting a visualization result (original image + probability distribution bar chart); finally, a classification report is generated, displaying the original input image for classification and its corresponding probability distribution and predicted labels for each category.

[0130] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for the detection of microplastics in wastewater by sieving, characterized in that, The method comprises: S1, collecting a sewage sample, separating and collecting a particle sample in the sewage sample, the particle sample being divided into microplastics and non-microplastics; S2, performing deep oxidation treatment on the collected particle sample to generate oxidation functional groups on the surface of the microplastics; S3, performing spectral detection on the particle sample after deep oxidation treatment by using a vibrational spectroscopy method, and selecting a particle sample with a spectrum matched with a polymer characteristic peak, a preset oxidation functional group characteristic peak existing in the spectrum, and a preset oxidation-body characteristic peak correlation rule being met; S4, marking the selected particle sample as microplastics.

2. The method of claim 1, wherein, The S2 specifically comprises: The particle sample is suspended in a hydrogen peroxide solution, and the particle sample in the solution is irradiated by ultraviolet rays; wherein the thickness of the suspension formed by suspending the particle sample in the hydrogen peroxide solution is not higher than 2 cm, the peak of the ultraviolet rays is 200 nm to 280 nm, the irradiation dose of the ultraviolet rays is 3600 mJ / cm 2 to 18000 mJ / cm 2 , and the mass concentration of the hydrogen peroxide solution is 1% to 10%. Alternatively, the particle sample is immersed in a potassium permanganate solution with a concentration of 0.01M-0.1M, and reacts at a temperature of 20-60°C for 10 minutes-6 hours.

3. The method of claim 1, wherein, The preset oxidation-body characteristic peak correlation rule comprises that the oxidation degree index of the particle sample is greater than a preset threshold value; The oxidation degree index is A 氧化峰 / A 参考峰 Wherein, A 参考峰 is the stable characteristic peak intensity or area of the bulk spectrum of the particle sample before deep oxidation treatment, A 氧化峰 is the peak area or peak height of the oxidation characteristic peak in the difference spectrum, and the difference spectrum is the difference between the spectrum after deep oxidation treatment and the spectrum before deep oxidation treatment.

4. The method of claim 1, wherein, The S2 and the S3 further comprise: The particle sample after deep oxidation treatment is transferred to a filter paper, and a positioning grid is overlaid on the filter paper; The filter paper is scanned by using an optical microscope, the position information of the particle sample in the positioning grid coordinate system is recorded, and a particle sample distribution image is generated; The S4 further comprises that, according to the marking result in the S4, label information corresponding to the microplastics and the non-microplastics in the particle sample distribution image is set.

5. The method of claim 4, wherein, The method further comprises: acquiring a plurality of particle sample distribution images with label information as a data set, training a preset sewage microplastic detection model by using the data set, and the sewage microplastic detection model is used for detecting the information of sewage microplastics in a target image.

6. The method of claim 5, wherein, The sewage microplastic detection model comprises a feature extraction enhancement module, a feature fusion enhancement module and a discrimination module connected in sequence; The feature extraction enhancement module comprises: a geometric morphological feature channel for extracting and enhancing the geometric morphological features of the target object in the image; a biological membrane feature channel for extracting and enhancing the local texture features generated by the target object being covered by a biological membrane; and a detail texture feature channel for extracting and enhancing the detail texture features representing the aging of the target object; The feature fusion enhancement module is used for fusing and enhancing the geometric morphological features, the local texture features and the detail texture features; The discrimination module is used for classifying and discriminating the fused and enhanced features to determine whether the target object is sewage microplastics.

7. The method of claim 6, wherein, The geometric morphological features comprise features of a fibrous structure, features of a fragment structure and features of a granular structure; and the image processing process of the geometric morphological feature channel comprises: The edge direction feature of the fibrous structure in the image is enhanced by using a direction gradient histogram feature map corresponding to the image, and the features of the fibrous structure are extracted by using a deep separable convolution; The features of the fragment structure in the image are adaptively extracted by using a deformable convolution network; A multi-scale ring convolution kernel is used to simulate a circular contour by using a weight distribution, so as to enhance the response to a circular closed boundary, thereby extracting the features of the granular structure in the image; and The features of the fragmented structure are enhanced based on a spatial attention mechanism; and the features of the granular structure are enhanced based on a radial maximum pooling operation.

8. The method of claim 7, wherein, The extracted features of the fibrous structure are as follows: ； in, The characteristics of the extracted fibrous structure, For dynamic convolution kernel weights, For local regions in the input feature map; The kernel size; As a learnable basic convolutional kernel, Here is the directional intensity function. Center pixel gradient direction, For the domain pixels gradient direction, To control the hyperparameters of orientation sensitivity, For the domain pixels The gradient magnitude.

9. The method of claim 6, wherein, The image processing process of the biofilm feature channel includes: Different levels of biofilm texture features are extracted by using several convolution layers with different expansion rates; After the biofilm texture features at different levels are fused, the biofilm texture enhanced features are obtained through feature enhancement; The biofilm texture enhanced features after adaptive average pooling and the biofilm texture enhanced features after adaptive maximum pooling are spliced to obtain spliced features; The features related to the microplastic material in the spliced features are enhanced by learning channel weights to form the local texture features.

10. The method of claim 6, wherein, The image processing process of the detail texture feature channel includes: Cascade maximum pooling is performed on the image to enhance the spatial sensitivity of holes and cracks in the image; The biofilm soft edge features and aging crack hard edge features in the image after cascade maximum pooling are detected by using a Sobel edge detection; the hole internal dark area features and long crack continuity features in the image after cascade maximum pooling are detected by using a hollow convolution; the local plaque texture features in the image after cascade maximum pooling are detected by using a point-by-point convolution; and the spatial attention map is generated by splicing the features of each branch; The spatial attention map and the local texture features are weighted and fused by using a learnable biofilm gain factor to obtain the detail texture features representing the aging of the target object.

Images