Pollen monitoring method and device
Through the automation technology combined with virtual impactors and electron microscopes, the problem of large labor and low accuracy of artificial microscopy is solved, efficient and accurate pollen monitoring and instant data release are achieved, and the efficiency and accuracy of pollen monitoring are improved. .
Patent Information
- Application Number
- CN202010011480.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-01-06
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2040-01-06
AI Technical Summary
The existing pollen monitoring methods rely on artificial microscopy, which have problems such as large labor, low accuracy, high influence by human factors, low efficiency, and the inability to achieve online monitoring and immediate forecasting.
The pollen particles in the air are collected by virtual impactors, and automatically classified and counted through electron microscopy and image analysis technology. The image processing is performed using multi-level classifiers and machine learning algorithms to fix the pollen particles to achieve repeated observations and data storage.
It greatly improves monitoring efficiency and accuracy, reduces labor intensity, realizes repeated observations of pollen particles and instant data release, and has online monitoring capabilities.
Smart Images

Figure CN111896430B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of environmental monitoring, in particular to a pollen monitoring method and a device thereof. Background Art
[0002] The current traditional collection method mainly uses gravity sedimentation to collect pollen, and then uses manual microscopy to classify and count pollen. The so-called manual microscopy method is that experienced observers carefully observe the pollen samples magnified 400 times under a biological optical microscope line by line, manually distinguish the types of pollen particles, and manually count the number of different pollen and the total number of pollen. The manual microscopy method has the following main disadvantages:
[0003] 1) The workload of observers is too heavy, which can easily cause physical fatigue and reduce the accuracy of monitoring;
[0004] 2) The monitoring staff has high observation skills and needs long-term and continuous technical training to accurately grasp the microscopic morphological characteristics of different pollen particles and correctly identify pollen. Once the staff is transferred, new observers must be retrained. This costs a lot of manpower and material resources;
[0005] 3) The monitoring results are affected by human factors, and the observation results obtained by different observers may vary greatly;
[0006] 4) During the peak period of pollen transmission, the manual resolution time will be prolonged, which may easily cause eye fatigue of the observer and affect the efficiency of pollen counting.
[0007] In addition, the shelf life of samples taken by the microscopy method is short, which makes it impossible to preserve the samples. If the data is lost, the samples cannot be recounted. In addition, the samples cannot be counted multiple times to prevent counting errors and other problems. Due to the above technical shortcomings, the current manual microscopy monitoring and analysis efficiency is not high, and it is difficult to guarantee the accuracy of its monitoring. In addition, because it is manual monitoring and does not have online capabilities, the data released during the peak period of pollen dispersal has an error of at least 24 hours from the peak period, and there is no way to make an immediate forecast to remind people to take precautions in advance. Summary of the invention
[0008] In view of the deficiencies of the prior art, the present invention provides a pollen monitoring method and also provides a monitoring device that can be used for the monitoring method.
[0009] The technical solution of the present invention is:
[0010] The pollen monitoring method comprises the following steps:
[0011] 1) Air sampling;
[0012] 2) Separate particles in the air according to size;
[0013] 3) Fixing the sampled particles on a glass slide;
[0014] 4) Microscope imaging of sampled particles;
[0015] 5) Image analysis (including classification and counting);
[0016] 6) Storing processed slides (containing analyzed particles);
[0017] 7) Carry out the next round of sampling and transmit the analysis results to the server according to the set time interval.
[0018] Furthermore, in step 1, a virtual impactor (also referred to as an impact sampler) is used to collect surrounding air and impact the air onto a glass slide coated with gelatin and glycerin, and pollen particles in the air are adhered by the glycerin, thereby achieving sampling.
[0019] Furthermore, in step 2, the air around the virtual impactor is heated to 60 m 3 / h to collect gas, and the virtual impactor takes the central airflow (about 6m 3 / h), the virtual impactor first separates particles of different sizes in the air and then pushes particles that match the size of pollen particles toward the slide.
[0020] Furthermore, in step 3, the gelatin is melted by heating, and the pollen particles originally adhered by glycerin are sunk into the gelatin, and then the pollen is permanently fixed in the gelatin by cooling, which is convenient for subsequent processes.
[0021] Furthermore, in step 4, the sample on the glass slide is scanned and photographed under an electron microscope to obtain an image stack and store it on a hard disk. Specifically:
[0022] Both the glass slide and the electron microscope can be rotated to a certain extent to facilitate the photography of samples at multiple angles. The electron microscope selects 120 positions for the sample on each glass slide to photograph, and uses image preprocessing (mainly including image transformation, image enhancement, edge detection and segmentation methods) to separate the shooting results according to the target particles and then re-stitch them into an image stack for subsequent analysis and processing and store them on the hard disk.
[0023] Furthermore, in step 5, the image stack obtained in step 4 is exported from the hard disk for image analysis, and the image stack is further classified, identified and counted for pollen through big data and machine learning algorithms. The process mainly includes the following:
[0024] 1) Feature extraction. Feature extraction is performed on the image stack of the target particles (mainly including pollen contour features, pollen structure features and pollen texture features). If a feature extraction fails, the particle is judged not to be pollen. If all the features are extracted successfully, the extracted parameter values constitute the feature vector set of each particle object (for the next step of automatic classification).
[0025] 2) Automatic (recognition) classification. The system uses a multi-level classifier (including three sub-classifiers: outer contour recognition module, membership-based internal structure recognition module, and texture recognition module) to classify the target pollen particles (determine the plant species corresponding to the pollen).
[0026] 3) Statistical counting: Repeat the above steps to process all electron microscope images taken within a period of time to obtain the number of pollens of different plant species.
[0027] The pollen monitoring device includes a rotating device, a rotating sample stage, a glass slide and a heating module. The rotating device is installed on a base. The rotating shaft of the rotating device is vertically arranged upward. The rotating shaft is connected to the rotating sample stage. The rotating sample stage is provided with a plurality of grooves. The grooves are circular and have through holes at the bottom. A glass slide is detachably placed in the grooves. A gelatin layer is arranged on the glass slide. A glycerin layer is arranged on the gelatin layer. A heating module is also installed on the base. The heating end of the heating module is arranged corresponding to the groove of the rotating sample stage, so that the glass slide can be heated by the heating end of the heating module when the groove rotates to the top of the heating module. When in use, pollen is impacted on the glass slide by an impact sampler or other equipment. At this time, the pollen is stuck by the glycerin layer, and then the rotating sample stage is rotated so that the glass slide in the groove is located above the heating module and heated. After heating, the pollen is trapped in the melted gelatin layer, and the pollen can be fixed after cooling. Then the rotating sample stage is rotated again, and the pollen sample is observed using an electron microscope.
[0028] In some embodiments, the rotating device adopts a graduated disk, so that the angle of each rotation can be precisely controlled.
[0029] In some embodiments, the rotating sample stage is made of a cast iron plate, a stainless steel plate or a high temperature resistant resin plate, which has sufficient hardness and is not afraid of high temperatures.
[0030] In some embodiments, the heating module is a far-infrared heater, and its heating direction is vertically upward. As is known, the far-infrared heater has limited heating effect on the air medium, so most of the heat can directly reach the glass slide, which improves the heating effect, reduces heat loss, and does not cause a significant increase in the temperature of the air around the device.
[0031] The beneficial effects of the present invention are: greatly improving monitoring efficiency, reducing labor intensity, improving monitoring accuracy, and also allowing samples to be repeatedly observed by fixing pollen particles. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a schematic diagram of the principle of the present invention;
[0033] Figure 2-3 This is a schematic diagram of the principle of step 4;
[0034] Figure 4 This is the schematic diagram of the multi-level classifier. DETAILED DESCRIPTION
[0035] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:
[0036] Example 1
[0037] like Figure 1-3 As shown, from the surrounding air at 60m 3 / h to collect gas, taking the central airflow (about 6m 3 / h) enters the virtual impactor for subsequent sampling and analysis. The virtual impactor first separates particles of different sizes in the air, and then pushes particles that meet the pollen particle size onto the glass slide (the glass slide is covered with gelatin and glycerol). The virtual impactor works continuously at the set interval until a sampling cycle ends (the sampling cycle can be set, such as 3 hours). The rotating sample stage rotates to the heating module, and the glass slide is heated to 90°C. The high temperature melts the sticky surface on the glass slide, causing the particles to sink into the medium, so as to fix and adjust the position of the particles, which is convenient for subsequent electron microscope observation and focusing. The rotating sample stage then rotates the glass slide to the position above the electron microscope, and then the sample is scanned and photographed under the electron microscope. Both the glass slide and the electron microscope can be rotated to a certain extent to facilitate the shooting of the sample at multiple angles. The shooting results are synthesized into a pollen image stack and stored on the hard disk. After the scan is completed, the loading and unloading system puts the sample slide into a storage box. This storage box has enough capacity to load a full month of samples.
[0038] The computer system loads the image stack from the hard disk for image analysis, separates pollen from other aerosol particles, and identifies pollen species based on their morphological characteristics.
[0039] The sampling period can be set by the operator, usually 1-6 hours / day, and each sampling period consumes one slide, which is stored in a reusable storage box.
[0040] The electron microscope selects 120 positions on each sample on the slide to take pictures. The pictures are separated according to the target particles through image preprocessing and form an image stack. The image stack is then analyzed by the graphic analysis module. Multiple automatic classification conditions are combined (including: length, width, radius, number of gaps, texture, etc.) during the analysis process. Finally, the image stack and the analysis results are stored together on the local hard disk.
[0041] The image analysis and processing flow is as follows:
[0042] 1) Image preprocessing: The electron microscope takes pictures of the sample from multiple angles, and uses image preprocessing (mainly including image transformation, image enhancement, edge detection and segmentation methods) to separate the shooting results according to the target particles and reassemble them into an image stack that is convenient for subsequent analysis and processing.
[0043] 2) Feature extraction. Feature extraction is performed on the image stack of the target particles (mainly including pollen contour features, pollen structure features and pollen texture features). If a feature extraction fails, the particle is judged not to be pollen. If all the features are extracted successfully, the extracted parameter values constitute the feature vector set of each particle object (for the next step of automatic classification).
[0044] 3) Automatic (recognition) classification. The system uses a multi-level classifier (including three sub-classifiers: outer contour recognition module, internal structure recognition module based on membership, and texture recognition module) to classify the target pollen particles (determine the plant species corresponding to the pollen).
[0045] 4) Statistical counting: Repeat the above steps to process all electron microscope images taken within a period of time to obtain the number of pollens of different plant species.
[0046] Remark:
[0047] When the automatic (identification) classification results cannot determine the pollen plant species, the image can be manually reviewed for determination at a later stage, and the determined plant species can be entered into the system. The system can then automatically identify pollen of that type when it encounters it.
[0048] The following describes in detail the key steps of image analysis and processing: image preprocessing, feature extraction, and automatic (recognition) classification:
[0049] 1. Image preprocessing
[0050] Image Transformation
[0051] The image transformation mentioned here refers to two-dimensional orthogonal transformation, which plays an important role in image processing. Image transformation is mainly considered for further image processing. For example, the average value after Fourier transformation is proportional to the average value of the image grayscale, and the high-frequency component indicates the edge information of the image. These properties can be used to extract feature values from the image. In this system, common image transformation methods such as Fourier transform, cosine transform, KL transform, wavelet transform, etc. are designed, and appropriate transformation methods are used according to different image features.
[0052] Image Enhancement
[0053] The purpose of image enhancement is to use a series of technologies to improve the visual effect of the image, increase the clarity of the image, or convert the image into a form that is more suitable for human observation and machine automatic analysis. For example, sharpening can highlight the edge contour of the image, and programming and controlling the computer to track it can perform various feature analyses. The methods for enhancing plant pollen images include image grayscale trimming, image smoothing, image sharpening, etc.
[0054] 1.3. Edge Detection and Segmentation
[0055] The edge of an image is the most basic feature of the image. The edge can outline the target object and make it clear to the observer. The edge contains rich intrinsic information such as the direction, step, shape, etc. of the image, and is an important attribute for extracting image features in image recognition. In essence, the edge of an image is a reflection of the discontinuity of the local characteristics of the image (grayscale mutation, color mutation, etc.), which marks the end of one area and the beginning of another. Edge extraction first detects the discontinuity of the local characteristics of the image, and then connects these discontinuous edge pixels into a complete boundary. The characteristic of the edge is that the pixels along the edge change slowly, while the pixels perpendicular to the edge change dramatically.
[0056] Image segmentation is a processing technology that divides an image into several meaningful regions. Here, "meaningful" generally refers to "corresponding to the target" or "a function of the problem being studied". For example, if the input is a terrain aerial photo or a landform remote sensing image, it is necessary to segment the mountains, plains, waters, forests, cities, roads, etc. When processing plant pollen images, it is necessary to detect the identification features of the pollen, such as holes and grooves. These "targets" separated from the image domain are the objects of image segmentation.
[0057] The basis of image segmentation is the similarity and jump between pixels. The so-called "similarity" means that the pixels in a certain area have similar characteristics, such as consistent grayscale and similar texture; the so-called "jump" refers to the discontinuity of characteristics, such as grayscale value mutation. There are many methods for image segmentation, which can be divided into threshold method, boundary detection method, matching method, tracking method, etc. This system uses edge extraction based on pixel grayscale mutation, region generation method based on characteristic similarity and texture analysis to study image segmentation.
[0058] 2. Feature extraction
[0059] In order to build a system to recognize different kinds of objects, we must first determine which characteristics of the object should be measured as descriptive parameters. These special properties that are measured are called features of the object, and the resulting parameter values constitute the feature vector of each object. It is very important to choose the features appropriately because it is the only basis for identifying the object.
[0060] In the image recognition problem, what needs to be extracted should be a small number of features with good distinguishability, reliability and independence. This system selects three features: the outline, structure and texture of pollen.
[0061] 2.1. Pollen contour feature extraction
[0062] Pollen contours are circular, elliptical, triangular, pincushion, and petal-shaped. Considering the large grayscale difference between pollen and its background, and the noise caused by small debris around the pollen, Gaussian filtering is used to first perform amplitude filtering to remove noise and make the obtained contour more regular. Edge tracking is performed after binarization. Edge tracking obtains a closed curve, and then the chain code of this curve is obtained. Based on this chain code, the characteristic values of the pollen contour, such as the perimeter, area, and regional roundness, are extracted, and the distance from the point on the closed curve to the center of mass is calculated to obtain another curve. The characteristics of this curve (such as how many peaks there are, whether it is close to a straight line, etc.) are used as the characteristics of the pollen contour.
[0063] 2.2. Pollen structural feature extraction
[0064] Since pollen of the same family or genus may have different outlines, and pollen of different families may have similar outlines, it is not enough to use only outline features. So the next step is to find structural features. The structure of pollen is mainly reflected in the internal pores and grooves. The grooves are thin and long, and the pores are round. Therefore, if the information of pores and grooves can be extracted and described, it will be an important feature. Since the pores and grooves appear concave in three dimensions, they appear as grayscale mutations under electron microscope. Therefore, the gradient operator is used to find its edges.
[0065] 2.3. Pollen texture feature extraction
[0066] Texture reflects certain changes in the color and grayscale of an object's surface, and these changes are related to the properties of the object itself. For example, wood of the same tree species has the same or similar textures, and people identify the species and material of wood by distinguishing the types of wood grain. The surface of pollen is smooth or wavy, and some pollen also has various carvings such as small thorns, tumors, particles, stripes, meshes, etc., forming a variety of textures. These textures are an important basis for identifying pollen species.
[0067] In image recognition, texture is a property that reflects the spatial distribution of the grayscale of pixels in a region. Therefore, when extracting texture features, the focus is on the texture measurement of an object in an image. If the grayscale level of an image is a constant or close to a constant, it means that the object has no texture. If the grayscale of an image changes significantly but is not a simple change in tone, then the object has texture. In order to measure texture, it is necessary to try to quantify the nature of the grayscale change of the image. The texture feature is a value calculated from the image of an object, which quantifies the characteristics of the grayscale change of the object image. First, the grayscale co-occurrence matrix is used to determine the coarseness and fineness of the texture. And the statistical parameters such as energy, entropy, contrast, and correlation are extracted from the grayscale co-occurrence matrix as texture feature parameters, which respectively reflect the uniformity of the grayscale distribution of the image, the amount of information in the image, the clarity of the image, and the similarity of the elements of the grayscale co-occurrence matrix in the row direction or column direction.
[0068] 3. Automatic (recognition) classification
[0069] 3.1. Principle of multi-class classifier
[0070] like Figure 4 As shown in the figure, the classifier can be regarded as a "machine" composed of hardware or software. Its function is to first calculate c discriminant functions gi, and then select the class corresponding to the maximum value of the discriminant function as the decision result.
[0071] A multi-level classifier is constructed in this system. The multi-level classifier applies different recognition techniques to pollen images with different characteristics, making full use of the advantages of different recognition methods, thus outperforming the single-level recognition system in performance.
[0072] In recent years, the combination of multiple classifiers has become a cutting-edge research topic in the field of pattern recognition, and has achieved good application results in many applications of pattern recognition, such as character recognition and target recognition. Many methods have been proposed for the combination of multiple classifiers. Commonly used methods include majority decision method, linear weighting method, Bayesian estimation, etc.
[0073] In view of the particularity of this system, a dynamic weight method for combining the results of multiple classifiers is designed. This method does not fix the weight of each classifier, but outputs a sequence of candidate sets and the corresponding credibility. The credibility of a candidate set refers to the reliability of the input pattern belonging to this pattern (or classification). If the credibility of classification result A in classifier x is much higher than other classification results, a higher weight is assigned to this classifier, that is, for a single classifier x, the higher the separation of the credibility of its classification results, the higher the weight of this classifier. The advantage of this algorithm is that when a single classifier can basically determine the classification result, this classifier dominates the combination of multiple classifiers. When the separation of the credibility of the classification results of all classifiers is similar, each classifier plays an approximate decision-making role.
[0074] 3.2. Multi-class classifier implementation
[0075] It consists of three sub-recognition modules (sub-classifiers):
[0076] (1) Outer contour recognition module
[0077] (2) Internal structure recognition module based on membership
[0078] (3) Texture recognition module
[0079] The correlation between classifiers at all levels can be divided into two cases: 1. Describes the same aspects of the same object, but to different degrees. 2. Describes different aspects of the same object. Case 1 reflects the redundant information of the classifier, and case 2 reflects the complementary information of the classifier. In the low-level information fusion - the fusion of sensor data, it is necessary to consider the situation where the influence of strong noise causes some sensors to fail. We can use redundant information to maintain the stability of the fusion system. In the high-level information fusion - the fusion of classifiers, redundant information will lead to redundant estimates of the combination results, so it needs to be removed.
[0080] Although complementary information describes different aspects of the same object, it is not always beneficial to the combination of classifiers. Only complementary information that can improve the combination effect is useful. The key to classifier combination is how to complement each classifier to make the combination better than each individual classifier. The division of the three sub-classifiers in this system is based on this consideration, maximizing complementarity and minimizing redundancy.
[0081] 3.2.1. Outer contour recognition module
[0082] The extraction of outer contour features includes the perimeter, area, regional roundness, number of corners, number of peaks of the centroid profile, and variance of the distance to the centroid. First, the sample images can be divided into two categories according to the regional roundness: circle and non-circle. When the outer edge is roughly judged to be a circle, the variance of the regional centroid profile curve can be used to determine whether the edge curve is regular. If it is a non-circle, the number of peaks of the regional centroid profile curve can roughly determine the number of edges of the pollen image. When the number of peaks is 2, there may be two cases: ellipse and pillow. Because they represent two major categories of pollen, it is necessary to distinguish them. Experiments have shown that the variance of the regional centroid profile curve has a good classification effect in distinguishing these two categories. If the variance is large, it can be considered as a pillow. If the variance is small, it is an ellipse. But the ellipse here is an ellipse with a large difference between the major axis and the minor axis, because the ellipse similar to the circle with similar major and minor axes has been classified as a circle. When the number of peaks is ≥3, it indicates that the outer contour shape of the pollen image is not easy to determine, and it should be considered in combination with the number of nodes on the edge. It can be simply considered that when the number of corner points is small, the edge is smoother, and the number of peaks is used as the number of sides of the outer contour; when the number of corner points is large, it is judged as an irregular shape.
[0083] The more difficult point is the determination of thresholds. Such as the area roundness, the variance of the area centroid profile curve and the number of corner points. In this system, the method used is mainly determined by empirical values. When these thresholds are determined, the classification interval is also determined. The distance between these thresholds and the characteristic values of the substitute samples determines the credibility of the substitute samples belonging to each possible classification.
[0084] 3.2.2. Internal structure recognition module based on membership
[0085] (1) Membership degree and membership function
[0086] A concept without a clear extension is called a fuzzy concept, and the entire object of discussion is called a domain or space. A fuzzy concept does not have an ordinary set as its extension, and the degree to which the elements on the domain conform to the concept is not absolutely 0 or 1, but can be between 0 and 1. In fuzzy mathematics, the absolute membership of elements to ordinary sets is made flexible, and the concept of membership is proposed. The membership degree can be described by a membership function.
[0087] A fuzzy set A on any domain S = {X} is a set of elements in x with certain properties whose boundaries are not clearly defined. It can be represented by the membership function F A (x) to characterize, F A The size of (x) reflects the degree of membership of x to the fuzzy set A, and its value is in the closed interval [0, 1]. A If the value of (x) is close to 1, it means that x is highly subordinate to A. AWhen the value of (x) is close to 0, it means that x is only slightly subordinate to A. For the domain, the elements of the domain are always distinct, and only the subsets A and B in S are fuzzy, so fuzzy sets usually refer to fuzzy subsets. In situations where confusion is not easy, fuzzy subsets are simply called fuzzy sets. For example, if S is a number domain, and the fuzzy set A represents a real number much greater than 0, that is, A = {x|x>>0}, then the membership function of A can be written as:
[0088]
[0089] (2) Internal structure classifier design
[0090] The method of directly determining the attribution of a sample by calculating its membership is called the membership principle of pattern classification, also known as the direct method of fuzzy pattern classification.
[0091] Affiliation Principles:
[0092] Assume that there are n fuzzy subsets in the domain U And each pair Membership function
[0093]
[0094] Then x0 is considered to belong to A i .
[0095] The membership principle is obvious and easily recognized, but its effectiveness depends heavily on the skill of establishing membership functions for known patterns.
[0096] The Gaussian membership function is used in this system. The main reason is that this function has the following good features: ① Simple expression, even for multiple input variables will not increase too much complexity. ② Radially symmetric, in line with the properties of ordinary objective things. ③ Positive definiteness.
[0097] The cluster center m of each category is obtained by cluster analysis i and standard deviation σ i , we can get the membership function of each class Extract the feature vector of the sample to be tested Substitute into the membership function of each class. According to the membership principle, the class with the maximum membership is the class to which the sample to be tested belongs. In this classifier, credibility can be simply considered as membership.
[0098] 3.2.3. Texture recognition module
[0099] (1) Texture feature selection by entropy minimization
[0100] Entropy is a statistical measure of uncertainty. For a given population of pattern vectors, the overall entropy is used to measure the intra-class dispersion, that is,
[0101] H=-E p {lnp(x)}
[0102] Where p(x) is the probability density of the model population, E p is the expected operation of p(x). When considering optimal feature selection, features that reduce uncertainty are considered to have more information. Therefore, when using entropy as an uncertainty measure, selecting features that minimize the entropy of the pattern class is a reasonable feature selection criterion.
[0103] Consider M pattern classes, whose probability densities are p(x|ω1), p(x|ω2), …, p(x|ω M ), then the entropy of the i-th pattern class is
[0104] H i =-∫ x p(x|ω i )ln[p(x|ω i )]dx
[0105] Its integration domain is the entire pattern space. Obviously, if p(x|ω i )=1, then H i = 0, there is no uncertainty. Therefore, the smaller the entropy, the smaller the uncertainty, and the better the separability.
[0106] In texture feature extraction, a variety of texture information features are obtained through gray level co-occurrence matrix and Fourier transform texture analysis method. Due to the large redundancy between features, the entropy of the entire pattern class is large and the separability is poor. In addition, the dimension of the feature vector is high, and the computational complexity is greatly increased. Therefore, it is necessary to reduce the dimension of the feature vector based on the principle of reducing entropy.
[0107] (2) Specific implementation
[0108] In this system, I first classify the same type of features into one category, and then obtain a feature quantity describing each type of feature by weighted average according to the separability of each feature. In this way, the redundancy between the same type of features is reduced. Experiments have shown that the entropy value of the pattern class is greatly reduced after feature selection compared to before feature selection.
[0109] The description of texture roughness includes: the angular second-order moment f1 extracted from the gray-level co-occurrence matrix; the gray-level difference probability P(d) and the entropy f4. r obtained by Fourier transform maxAccording to the value of each feature, the roughness expressed by each feature is estimated first, and then the weighted average is taken to obtain the uniform roughness C0. max The separability is better, so it is assigned a larger weight. The grayscale difference probability P(d) entropy f4 is assigned a smaller weight.
[0110] Because the same type of pollen images may rotate in different directions, this system focuses on extracting the strength of the pollen texture direction rather than the specific direction when extracting texture direction features. The system uses three features: MAX P i (r) indicates the ring characteristic curve P i (r) Whether the peak is obvious, the number of peaks N and the directional feature DIR extracted from the power spectrum matrix. Mainly composed of MAX P i (r) determines, N and DIR play an auxiliary role. Get the directional strength I d .
[0111] The experiment found that some other texture features also played a good role in pollen texture classification:
[0112] Contrast of Gray Level Co-occurrence Matrix Reflects the clarity of the texture.
[0113] Correlation of gray-level co-occurrence matrix It reflects the similarity of the elements of the gray level co-occurrence matrix in the row direction or column direction.
[0114] After selecting the above features, the average of the features of each type of pollen in the pollen image library is calculated, the feature value of the sample to be tested is compared with the average, and the sample to be tested is classified into the pollen class with the smallest distance. The distance is standardized and used as the credibility of each classification.
[0115] Example 2
[0116] A pollen monitoring device is used, which includes a rotating device, a rotating sample stage, a glass slide and a heating module. The rotating device is installed on a base, and the rotating shaft of the rotating device is vertically arranged upward. The top of the rotating shaft is connected to the rotating sample stage. A plurality of grooves are arranged on the rotating sample stage. The grooves are circular and have through holes at the bottom. A glass slide is detachably placed in the grooves. A gelatin layer is arranged on the glass slide, and a glycerin layer is arranged on the gelatin layer. A heating module is also installed on the base. The heating end of the heating module is arranged corresponding to the groove of the rotating sample stage, so that the glass slide can be heated by the heating end of the heating module when the groove rotates to the top of the heating module. When in use, pollen is impacted on the glass slide by an impact sampler or other equipment. At this time, the pollen is stuck by the glycerin layer, and then the rotating sample stage is rotated so that the glass slide in the groove is located above the heating module and heated. After heating, the pollen is trapped in the melted gelatin layer, and the fixation of the pollen can be achieved after cooling. Then the rotating sample stage is rotated again, and the pollen sample is observed using an electron microscope.
[0117] The rotating device adopts a dividing plate, and the angle of each rotation can be accurately controlled.
[0118] The rotating sample stage is made of a cast iron plate, a stainless steel plate or a high temperature resistant resin plate, which has sufficient hardness and is not afraid of high temperatures.
[0119] The heating module is a far-infrared heater, and its heating direction is vertically upward. As is known, the far-infrared heater has limited heating effect on the air medium, so most of the heat can directly reach the slide, which improves the heating effect, reduces heat loss, and does not cause a significant increase in the temperature of the air around the device.
[0120] One way to use it:
[0121] Through the impact collector from the surrounding air at 60m 3 / h speed to collect gas so that pollen will hit the glass slide (of course, there may be no pollen in the air, then there will be no pollen on the glass slide). The collection cycle should be fixed, for example, once every 3 hours.
[0122] Rotate the rotating sample stage to the heating module, heat the slide to 90°C, and the high temperature melts the sticky surface on the slide so that the particles sink into the medium, so as to fix and adjust the position of the particles, which is convenient for subsequent electron microscope observation and focusing. The rotation can be manually controlled or automatically controlled by control devices such as PLC (installed by the user). The heating temperature requires the user to obtain the best heating module parameters through a limited number of experiments.
[0123] After heating, the rotating sample stage is rotated to make the heated glass slide reach the observation position, at which time it can be observed and recorded through an electron microscope. After the recording is completed, the glass slide with the pollen sample fixed can be taken out and placed in a storage box. Then, the graphic stack analysis is performed to achieve pollen classification.
[0124] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.
Claims
1. Pollen monitoring methods, It is characterized in that The following steps are involved: 1) Air sampling; 2) Separate particles in the air according to size; 3) Fixing the sampled particles on a glass slide; 4) Microscope imaging of sampled particles; 5) Image analysis; 6) Storing processed slides; 7) Carry out the next round of sampling and transmit the analysis results to the server according to the set time interval; In the step 1, a virtual impactor is used to collect surrounding air and impact the air onto a glass slide, the glass slide is coated with gelatin and glycerin, and pollen particles in the air are adhered by the glycerin, thereby achieving sampling; In step 2, the virtual impactor collects gas from the surrounding air at a rate of 60 m³ / h. The virtual impactor takes the central airflow, first separates particles of different sizes in the air, and then pushes particles that meet the pollen particle size toward the slide; In step 3, the gelatin is melted by heating, and the pollen particles originally adhered by the glycerin are sunk into the gelatin, and then the pollen is permanently fixed in the gelatin by cooling; Scan and photograph the sample on the slide using an electron microscope, obtain an image stack and store it on a hard drive; Both the glass slide and the electron microscope can be rotated to a certain extent to facilitate the shooting of samples at multiple angles. The electron microscope selects 120 positions of the sample on each glass slide to shoot, and uses image preprocessing to re-separate the shooting results according to the target particles and splice them into image stacks, which are then stored on the hard disk; In the step 5, the image stack obtained in the step 4 is exported from the hard disk for image analysis, and the image stack is further classified and identified for pollen through big data and machine learning algorithms, and on this basis, the classification statistics and counting of pollen are realized; The method is implemented using a pollen monitoring device, which includes a rotating device, a rotating sample stage, a glass slide and a heating module. The rotating device is installed on a base, and the rotating shaft of the rotating device is vertically upward. The top of the rotating shaft is connected to the rotating sample stage. A plurality of grooves are arranged on the rotating sample stage. The grooves are circular and have through holes at the bottom. A glass slide is detachably placed in the grooves. A gelatin layer is arranged on the glass slide, and a glycerol layer is arranged on the gelatin layer. A heating module is also installed on the base, and the heating end of the heating module is arranged corresponding to the groove of the rotating sample stage, so that the glass slide can be heated by the heating end of the heating module when the groove rotates to the top of the heating module.
