Dust pollution evaluation method based on deep residual network

Abstract

The embodiment of the present disclosure provides a dust pollution evaluation method based on a deep residual network, and belongs to the technical field of image recognition, and specifically comprises: forming an initial image set; performing a preprocessing operation on each group of images in the initial image set respectively to obtain a binary distribution image set; calculating the gray mean value of each group of images in the binary distribution image set respectively; screening the gray mean value result set according to a statistical standard and eliminating abnormal data; obtaining a linear function relationship between the gray mean value and the dust quality; obtaining an initial image grade division set through the corresponding relationship between the gray mean value and the initial image set; performing expansion processing on the image grade division set to obtain an image grade expansion set; inputting the image grade expansion set as training data into a preset image classification convolutional neural network model, and adjusting parameters to obtain a dust pollution grade evaluation model for rating a to-be-detected image. Through the scheme of the present disclosure, the detection efficiency, accuracy and adaptability are improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of image recognition technology, and in particular to a dust pollution evaluation method based on a deep residual network. Background Art

[0002] At present, in underground mines, dust pollution poses a threat to human health and property safety. It is now receiving increasing attention from mining companies. Real-time and accurate dust monitoring is also rapidly becoming the key to dust pollution control.

[0003] Existing methods for monitoring dust pollution in underground mines can be categorized as either offline or online. Offline methods typically use manual sampling equipment, such as membrane weighing devices and photoelectric rapid dust measurement devices. Offline methods suffer from limitations such as difficult to guarantee data accuracy, limited measurement areas, and long processing cycles. Online methods primarily rely on various types of sensors, which also come with significant drawbacks, such as fragile communication lines, difficulty controlling accuracy, and expensive maintenance.

[0004] It can be seen that there is an urgent need for a dust pollution evaluation method based on deep residual network with high detection efficiency, accuracy and adaptability. Summary of the Invention

[0005] In view of this, the embodiments of the present disclosure provide a dust pollution evaluation method based on a deep residual network, which at least partially solves the problems of poor detection efficiency, accuracy and adaptability in the existing technology.

[0006] The present disclosure provides a dust pollution evaluation method based on a deep residual network, including:

[0007] Step 1: Collect multiple sets of dust images after uniform diffusion of dust of different masses and arrange them in order to form an initial image set;

[0008] Step 2: Preprocess each group of images in the initial image set to obtain a binary image with each pixel value being 1 or 0. Pixel points with pixel values ​​of 1 and 0 represent dust particles and gaps between particles or other background elements, respectively, to obtain a binary distribution image set.

[0009] Step 3, respectively calculating the grayscale mean of each group of images in the binary distribution image set to obtain a grayscale mean result set;

[0010] Step 4: Screen the grayscale mean result set according to statistical standards and remove abnormal data to obtain the grayscale mean determination set;

[0011] Step 5, the arithmetic mean of the gray mean value determination set is calculated to obtain the gray mean value arithmetic mean set, and the gray mean value arithmetic mean set is linearly fitted with the dust mass to obtain a linear function relationship between the gray mean value and the dust mass;

[0012] Step 6, the measured dust mass interval is equally divided into T dust mass level intervals, and the corresponding pollution levels of the dust mass level intervals are determined in order from small to large, and then the gray mean value level interval is calculated through the linear function relationship between the gray mean value and the dust mass to correspond to the pollution level, to obtain T sets of gray mean value level division sets, and the initial image level division set is obtained through the corresponding relationship between the gray mean value and the initial image set;

[0013] Step 7, the image level division set is expanded to obtain an image level expansion set;

[0014] Step 8, the image level expansion set is input into a preset image classification convolutional neural network model as training data, and the parameters of the model are adjusted to obtain a dust pollution level evaluation model.

[0015] According to a specific implementation manner of an embodiment of the present disclosure, the preprocessing operation includes obtaining image gradient, gray scale, sharpening, binarization and open operation denoising.

[0016] According to a specific implementation manner of an embodiment of the present disclosure, the expansion processing includes at least one of flipping, scaling, rotating, gray scale, sharpening or adjusting brightness.

[0017] According to a specific implementation manner of an embodiment of the present disclosure, the calculation formula of the gray mean value is

[0018]

[0019] wherein, is the gray mean value, Z1 is the number of pixels with a gray value of 1 in the binary image, and M and N are the height and width of the binary image input size respectively.

[0020] According to a specific implementation manner of an embodiment of the present disclosure, the step 4 specifically includes:

[0021] The Chauvenet criterion is used to judge whether the residual absolute value of the data in the gray mean value result set meets the discriminant, the data of the check anomaly is deleted from the gray mean value result set, so that the gray mean value in the retention set is all valid data and is used as the gray mean value determination set, wherein the discriminant of the Chauvenet criterion is

[0022]

[0023] wherein, x iis the calculated grayscale mean data, Z c is the coefficient that varies according to the number of trials n, and σ(X) are the arithmetic mean and standard deviation estimates of the grayscale mean of images under the same quality conditions, respectively.

[0024] According to a specific implementation of the embodiment of the present disclosure, step 5 specifically includes:

[0025] In the grayscale mean determination set, the grayscale means of the same order between groups are selected, and the arithmetic mean of the grayscale means corresponding to the same test mass is calculated. The one-to-one correspondence between dust mass and arithmetic mean is expressed as the following function:

[0026]

[0027] in, Represents the arithmetic mean of the grayscale mean of group order n, m n Indicates the mass of the dust test with group order n;

[0028] Then, a linear fitting was performed on the relationship between the grayscale mean arithmetic mean and the dust mass to obtain the linear function relationship between the grayscale mean and the dust mass:

[0029]

[0030] Where a and b represent the slope and intercept of the linear relationship, respectively.

[0031] According to a specific implementation of the embodiment of the present disclosure, step 6 specifically includes:

[0032] Calculate the mass value that divides the measured dust mass interval into equidistant intervals, divide the measured dust mass interval into T dust mass grade intervals, which correspond to pollution levels 1 to T from left to right, and obtain the coordinates of the dividing points through the linear relationship between the grayscale mean and the dust mass. Use the vertical coordinate as the grayscale mean that divides the grayscale mean grade interval, and obtain T grayscale mean grade intervals, which correspond to pollution levels 1 to T from left to right.

[0033] According to a specific implementation of the embodiment of the present disclosure, before step 8, the method further includes:

[0034] Construct a convolution layer with a convolution kernel size of fc×fc×Cc;

[0035] Construct an activation layer after the convolutional layer using ReLU as the activation function;

[0036] Construct a pooling layer with a pooling window size of fp×fp, including an average pooling layer and a maximum pooling layer;

[0037] After the pooling layer, an activation layer is constructed using ReLU as the activation function;

[0038] Three types of residual modules are constructed through convolutional layers and pooling layers, each containing x convolutional layers and y pooling layers;

[0039] Construct a softmax layer;

[0040] Construct two fully connected layers to form a preset image classification convolutional neural network model.

[0041] According to a specific implementation of the embodiment of the present disclosure, step 8 specifically includes:

[0042] The image level expansion set is input into the preset image classification convolutional neural network model. The weight parameters W and bias parameters b of all matrix operations in the neural network are updated using the loss function L of the internal network layer and the gradient descent optimizer. After repeated iterations, the dust pollution level evaluation model is obtained.

[0043] The image to be detected is input into the dust pollution level evaluation model to obtain the pollution rating result.

[0044] According to a specific implementation of the embodiment of the present disclosure, the expression of the loss function is:

[0045]

[0046]

[0047] Where N is the total number of image samples, L ij is the loss function of a single image sample, y it is the screening function for correct prediction of image samples, p(i, j, t) is the pollution level prediction probability of the image sample, z(i, j, t) represents the pollution level prediction weight of the image sample obtained by neural network operation, i represents the pollution level of the level expansion set where the image sample is located, j represents the sequence number of the image sample in the level expansion set, t represents the pollution level predicted for the sample, and T is the number of pollution level divisions.

[0048] The dust pollution evaluation scheme based on deep residual network in the embodiment of the present disclosure includes: step 1, collecting multiple groups of dust images after uniform diffusion of dust of different masses and arranging them in order to form an initial image set; step 2, performing preprocessing operations on each group of images in the initial image set respectively to obtain a binary image with each pixel value of 1 or 0, using pixel points with pixel values ​​of 1 and 0 to represent dust particles and gaps between particles or other background elements, respectively, to obtain a binary distribution image set; step 3, calculating the grayscale mean of each group of images in the binary distribution image set respectively to obtain a grayscale mean result set; step 4, screening the grayscale mean result set according to statistical standards and eliminating abnormal data to obtain a grayscale mean determination set; step 5, performing arithmetic mean calculation on the grayscale mean determination set to obtain a grayscale mean arithmetic mean set, and comparing it with the dust mass The linear fitting is performed on the grayscale mean to obtain a linear function relationship between the grayscale mean and the dust mass; step 6, the measured dust mass interval is equally divided into T dust quality grade intervals, and the corresponding pollution levels of the dust quality grade intervals are determined in order from small to large quality, and then the grayscale mean grade intervals are calculated by the linear function relationship between the grayscale mean and the dust mass and correspond to the pollution level to obtain T groups of grayscale mean grade division sets, and the initial image grade division set is obtained by the correspondence between the grayscale mean and the initial image set; step 7, the image grade division set is expanded to obtain an image grade expansion set; step 8, the image grade expansion set is input as training data into a preset image classification convolutional neural network model, and the parameters of the model are adjusted to obtain a dust pollution grade evaluation model for rating the image to be detected.

[0049] The beneficial effects of the embodiments of the present disclosure are: through the scheme of the present disclosure, a connection between dust images and dust concentration is established, making dust concentration measurement more efficient and convenient. At the same time, the residual neural network can also train new image samples, greatly improving the detection efficiency, accuracy and adaptability of dust pollution evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0051] Figure 1 A schematic diagram of a process for evaluating dust pollution based on a deep residual network according to an embodiment of the present disclosure;

[0052] Figure 2 A schematic diagram of a physical device usage flow corresponding to a dust pollution assessment method based on a deep residual network provided in an embodiment of the present disclosure;

[0053] Figure 3 A schematic diagram of a dust image processing process provided by an embodiment of the present disclosure;

[0054] Figure 4 A schematic diagram of grayscale mean distribution provided in an embodiment of the present disclosure;

[0055] Figure 5 A grayscale mean value level interval division diagram provided in an embodiment of the present disclosure;

[0056] Figure 6 A schematic diagram of the accuracy change process of a training and verification process provided by an embodiment of the present disclosure;

[0057] Figure 7 A schematic diagram of a test accuracy result provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0058] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0059] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0060] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0061] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0062] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0063] The embodiments of the present disclosure provide a dust pollution evaluation method based on a deep residual network, which can be applied to dust pollution monitoring processes in scenes such as construction or underground mines.

[0064] See also Figure 1 , is a flow chart of a dust pollution evaluation method based on a deep residual network provided by an embodiment of the present disclosure. Figure 1 As shown, the method mainly includes the following steps:

[0065] Step 1: Collect multiple sets of dust images after uniform diffusion of dust of different masses and arrange them in order to form an initial image set;

[0066] In specific implementation, a group of images of dust with different masses evenly diffused can be tested and collected and arranged in order to obtain a group of initial image sets PGroup(1,2…a). The i-th group of initial image sets can be expressed as {P i1 ,P i2 …P in Specifically, the dust mass range for the test is 0-500 mg. A dust sample is prepared at every 5 mg interval, and arranged in descending order as 5 mg, 10 mg, 15 mg...500 mg. The preparation is repeated three times. Then, these dust samples are uniformly diffused in the test device in turn. Three sets of initial image sets are obtained by the digital camera in the test device. Each set of initial image sets contains 100 initial images. The test device and process are as follows: Figure 2 shown.

[0067] Step 2: Preprocess each group of images in the initial image set to obtain a binary image with each pixel value being 1 or 0. Pixel points with pixel values ​​of 1 and 0 represent dust particles and gaps between particles or other background elements, respectively, to obtain a binary distribution image set.

[0068] Optionally, the preprocessing operations include obtaining image gradients, grayscale conversion, sharpening, binarization, and opening operation denoising.

[0069] In specific implementation, the dust image collected in step 1 can be processed into a binary image with each pixel value being 1 or 0, and the pixel points with pixel values ​​1 and 0 are used to represent dust particles and the gaps between particles or other background elements, respectively, to obtain a group of dust binary distribution image sets BGroup(1,2…a). The i-th group of binary distribution image sets can be expressed as {B i1 ,B i2 …B in}; The specific image processing process is as follows Figure 3 As shown,

[0070] First, calculate the image gradient. The calculation formula is as follows:

[0071]

[0072] Where G k (x,y) is the image gradient, g(x) is the horizontal component of the image gradient, g(y) is the vertical component of the image gradient, f(x,y) is the pixel value of the image, x and y are the horizontal and vertical coordinates of the pixel in the image respectively;

[0073] Then the initial image is grayscaled. The grayscale formula is as follows:

[0074] Gr(x,y)=R*0.299+G*0.587+B*0.114

[0075] Where Gr(x,y) is the grayscale value of the pixel in the grayscale image, R, G, and B are the grayscale values ​​of red, green, and blue in the original image respectively;

[0076] Next, the grayscale image is sharpened to enhance the high-frequency part of the image and retain the details of the edges and contours within the image; then the image is normalized. The normalization formula is as follows:

[0077]

[0078] In the formula, Gr min and Gr max is the minimum and maximum grayscale values ​​of the image before normalization, E min and E max is the minimum and maximum grayscale values ​​of the image after normalization, and E(x,y) is the grayscale value of the image after normalization;

[0079] Then, binarization is performed, and the grayscale value of pixels whose grayscale value is greater than the adaptive segmentation threshold is assigned to 1, and the grayscale value of pixels whose grayscale value is less than the adaptive segmentation threshold is assigned to 1. The calculation formula of the adaptive segmentation threshold is:

[0080]

[0081] Where T is the adaptive segmentation threshold, W k (x,y) is the Gaussian weight, I k (x,y) is the grayscale value of the pixel, C is a constant;

[0082] At the end of image processing, the opening operation is performed to remove noise. The formula for the opening operation is:

[0083]

[0084] Where A is the input image, B is the structural element, and The symbols for erosion and dilation calculations, respectively.

[0085] Step 3, respectively calculating the grayscale mean of each group of images in the binary distribution image set to obtain a grayscale mean result set;

[0086] Furthermore, the calculation formula of the grayscale mean is:

[0087]

[0088] in, is the grayscale mean, Z1 is the number of pixels with grayscale value 1 in the binary image, and M and N are the height and width of the binary image input size respectively.

[0089] In specific implementation, the grayscale mean can be calculated for the processed image obtained in step 2 to obtain a grayscale mean result set VGroup(1,2…a) corresponding to the image. The grayscale mean result set of group i can be expressed as Specifically, the grayscale mean of each image is calculated separately to obtain three sets of grayscale mean values. The grayscale mean distribution is as follows: Figure 4 As shown in Figure 2, the calculation formula for the grayscale mean is:

[0090]

[0091] Where, is the grayscale mean, Z1 is the number of pixels with grayscale value 1 in the binary image, and M and N are the height and width of the binary image input size respectively.

[0092] Step 4: Screen the grayscale mean result set according to statistical standards and remove abnormal data to obtain the grayscale mean determination set;

[0093] Based on the above embodiment, step 4 specifically includes:

[0094] The Chauvenet criterion is used to determine whether the absolute value of the residual of the data in the gray mean result set satisfies the discriminant. The abnormal data are deleted from the gray mean result set, so that the gray mean retained set is all valid data and is used as the gray mean determination set. The discriminant of the Chauvenet criterion is:

[0095]

[0096] Among them, x i is the calculated grayscale mean data, Z c is the coefficient that varies according to the number of trials n, and σ(X) are the arithmetic mean and standard deviation estimates of the grayscale mean of images under the same quality conditions, respectively.

[0097] In the specific implementation, according to the statistical standard, the grayscale mean result set obtained in step 3 is screened and the corresponding abnormal data is discarded to obtain a group of grayscale mean determination set RGroup(1,2…a); specifically, the Chauvenet criterion is first used to determine whether the residual absolute value of the data in the grayscale mean result set meets the discriminant, and the data with abnormal verification is deleted from the grayscale mean result set VGroup(1,2…a), so that the grayscale mean retention set RGroup(1,2…a) contains valid data. The discriminant of the above Chauvenet criterion is:

[0098]

[0099] Where x i is the calculated grayscale mean data, Z c is the coefficient that varies according to the number of trials n, and σ(X) are the arithmetic mean and standard deviation estimates of the grayscale mean of images under the same quality conditions, respectively.

[0100] Step 5: Perform arithmetic mean calculation on the grayscale mean determination set to obtain the grayscale mean arithmetic mean set, and perform linear fitting on the grayscale mean set with the dust mass to obtain a linear function relationship between the grayscale mean and the dust mass;

[0101] Furthermore, the step 5 specifically includes:

[0102] In the grayscale mean determination set, the grayscale means of the same order between groups are selected, and the arithmetic mean of the grayscale means corresponding to the same test mass is calculated. The one-to-one correspondence between dust mass and arithmetic mean is expressed as the following function:

[0103]

[0104] in, Represents the arithmetic mean of the grayscale mean of group order n, mn Indicates the mass of the dust test with group order n;

[0105] Then, a linear fitting was performed on the relationship between the grayscale mean arithmetic mean and the dust mass to obtain the linear function relationship between the grayscale mean and the dust mass:

[0106]

[0107] Where a and b represent the slope and intercept of the linear relationship, respectively.

[0108] In specific implementation, the grayscale mean determination set of group a obtained in step 4 can be calculated by arithmetic mean to obtain the grayscale mean arithmetic mean set Then, the obtained arithmetic mean set is linearly fitted with the dust mass to obtain the linear function relationship between the grayscale mean and the dust mass; specifically, in the three sets of grayscale mean effective value sets, the grayscale means corresponding to the same dust test mass are calculated by arithmetic mean, and a set of arithmetic mean values ​​of the grayscale means is obtained, in which the arithmetic mean values ​​of the grayscale means correspond to the dust sample mass one by one, and then a linear fitting is performed to obtain the linear relationship between the grayscale mean and the dust mass.

[0109] Step 6: Divide the measured dust mass interval into T dust mass grade intervals with equal spacing. Determine the pollution levels corresponding to the dust mass grade intervals in ascending order of quality. Then, calculate the grayscale mean grade intervals using the linear function relationship between the grayscale mean and the dust mass and correspond them to the pollution levels. Obtain T groups of grayscale mean grade partition sets. Obtain the initial image grade partition set using the correspondence between the grayscale mean and the initial image set.

[0110] Based on the above embodiment, step 6 specifically includes:

[0111] Calculate the mass value that divides the measured dust mass interval into equidistant intervals, divide the measured dust mass interval into T dust mass grade intervals, which correspond to pollution levels 1 to T from left to right, and obtain the coordinates of the dividing points through the linear relationship between the grayscale mean and the dust mass. Use the vertical coordinate as the grayscale mean that divides the grayscale mean grade interval, and obtain T grayscale mean grade intervals, which correspond to pollution levels 1 to T from left to right.

[0112] In specific implementation, the measured dust mass interval can be equally divided into T dust mass grade intervals, and the corresponding pollution levels of the dust mass grade intervals are determined in order from small to large. Then, the linear relationship between the grayscale mean and the dust mass obtained in step 5 is used to calculate the grayscale mean grade interval and correspond to the pollution level, and T groups of grayscale mean grade division sets are obtained, among which the kth group of grayscale mean grade division sets can be expressed as Finally, the corresponding relationship between the grayscale mean and the initial image is used to obtain the T-group initial image classification set PLeve l(1,2…T); specifically, according to the test range of dust mass 0-500mg, four levels are divided, level I is 0-125mg, level II is 125-250mg, level III is 250-375mg, and level IV is above 375mg. The level interval corresponding to the grayscale mean is obtained through the linear relationship between the grayscale mean and the dust mass. The level interval of the grayscale mean is as follows: Figure 5 As shown, level I is 0-0.102, level II is 0.102-0.1395, level III is 0.1395-0.177mg, and level IV is above 0.177. Finally, all the initial images are classified according to the grayscale mean value corresponding to each image, and four image sets corresponding to different levels are obtained.

[0113] For example, we can first calculate the dust mass interval (0, M T ]Equally spaced quality values ​​M1, M2, M3…M T-1 , change (0,M T ] is divided into (0,M1], (M1,M2], (M2,M3]…(M T-1 ,M T ] Such T dust quality level intervals correspond to pollution levels 1-T from left to right, and then the segmentation point coordinates are obtained through the linear relationship between the grayscale mean and the dust quality. The vertical coordinate As the grayscale mean value for dividing the grayscale mean level interval, T grayscale mean level intervals are obtained. From left to right, they correspond to pollution levels 1 to T. Then the grayscale mean level partition set VLeve l(t) with pollution level t can be expressed as The initial image level partition set PLeve l(t) with pollution level t is in Represents the function between the initial image and the grayscale mean, which can be listed as:

[0114]

[0115] Where, P ij is the image with order j in the i-th group of initial images, is the grayscale mean of the i-th group of grayscale mean results with order j

[0116] Step 7: Expand the image level classification set to obtain an image level expansion set;

[0117] Optionally, the expansion processing includes at least one of flipping, scaling, rotating, graying, sharpening or adjusting brightness.

[0118] In particular implementation, the image level division set obtained in step 6 can be expanded by at least one of flipping, scaling, rotating, graying, sharpening, or adjusting brightness to obtain an image level expansion set ELevel(1, 2…T); specifically, 900 image samples are added by horizontal, vertical and diagonal flipping, 1200 samples are added by scaling by 0.25, 0.5, 2 and 4 times, 900 samples are added by rotating by 90°, 180° and 270°, 300 samples are added by graying, 300 samples are added by sharpening, 900 samples are added by adjusting brightness by 0.5, 1.5 and 2 times, and a total of 4800 image samples are added to the initial image samples.

[0119] In step 8, the image level expansion set is input into a preset image classification convolutional neural network model, and the parameters of the model are adjusted to obtain a dust pollution level evaluation model.

[0120] Optionally, before step 8, the method further comprises:

[0121] A convolution layer with a convolution kernel size of fc×fc×Cc is constructed;

[0122] An activation layer is constructed after the convolution layer with ReLU as the activation function;

[0123] A pooling layer with a pooling window size of fp×fp is constructed, including an average pooling layer and a maximum pooling layer;

[0124] An activation layer is constructed after the pooling layer with ReLU as the activation function;

[0125] Three residual modules are constructed through the convolution layer and the pooling layer, each containing x convolution layers and y pooling layers;

[0126] A softmax layer is constructed;

[0127] Two fully connected layers are constructed to form the preset image classification convolutional neural network model.

[0128] Further, step 8 specifically comprises:

[0129] The image level expansion set is input into the preset image classification convolutional neural network model, the loss function L and the gradient descent optimizer of the internal network layer are used to update the weight parameters W and the bias parameters b of all matrix operations in the neural network, and the dust pollution level evaluation model is obtained after repeated iterations;

[0130] The image to be detected is input into the dust pollution level evaluation model to obtain the pollution rating result.

[0131] Furthermore, the loss function is expressed as

[0132]

[0133]

[0134] Where N is the total number of image samples, L ij is the loss function of a single image sample, y it is the screening function for correct prediction of image samples, p(i, j, t) is the pollution level prediction probability of the image sample, z(i, j, t) represents the pollution level prediction weight of the image sample obtained by neural network operation, i represents the pollution level of the level expansion set where the image sample is located, j represents the sequence number of the image sample in the level expansion set, t represents the pollution level predicted for the sample, and T is the number of pollution level divisions.

[0135] In specific implementation, an image classification convolutional neural network model can be constructed, and the image level expansion set obtained in step S7 can be input into the model as training data, and the parameters of the model can be adjusted to finally obtain a comprehensive evaluation system for dust pollution levels; specifically, a neural network is constructed through deep learning technology, and 80% of the image samples are randomly extracted and input into the neural learning network for calculation. During the calculation process, a method of storing feature vectors by creating a vector folder is used to avoid repeated calculations until the network reaches convergence and completes training. Then, 10% of the image samples are randomly extracted for verification accuracy, and the remaining 10% of the image samples are used to test the accuracy of the evaluation system. The accuracy of the training and verification process is as follows: Figure 6 As shown, the test accuracy results are as follows Figure 7 As shown, the test accuracy finally reached 89.2%;

[0136] The specific method of constructing the image classification convolutional neural network is:

[0137] a constructs a convolution kernel of size f c ×f c ×C c Convolutional layers;

[0138] b. Construct an activation layer after the convolutional layer using ReLU as the activation function;

[0139] c builds a pooling window of size f p ×f p Pooling layers, including average pooling layer and maximum pooling layer;

[0140] d. Construct an activation layer after the pooling layer using ReLU as the activation function;

[0141] e constructs three types of residual modules through convolutional layers and pooling layers, which contain x convolutional layers and y pooling layers respectively.

[0142] f builds the softmax layer, there is 1 in total;

[0143] g builds a fully connected layer, with a total of 2;

[0144] The specific method for adjusting network parameters during model training is:

[0145] The image level expansion set is input into the constructed model, and the loss function L of the internal network layer and the gradient descent optimizer are applied to update the weight parameters W and bias parameters b of all matrix operations in the neural network. After repeated iterations, the ideal parameters are obtained.

[0146] The selected loss function is defined as follows:

[0147]

[0148]

[0149] Where N is the total number of image samples, L ij is the loss function of a single image sample, y it is the screening function for correct prediction of image samples, p(i, j, t) is the pollution level prediction probability of the image sample, z(i, j, t) represents the pollution level prediction weight of the image sample obtained by neural network operation, i represents the pollution level of the level expansion set where the image sample is located, j represents the sequence number of the image sample in the level expansion set, t represents the pollution level predicted for the sample, and T is the number of pollution level divisions.

[0150] The dust pollution evaluation method based on deep residual network provided in this embodiment establishes a connection between dust images and dust concentration, making dust concentration measurement more efficient and convenient. At the same time, the residual neural network can also be trained on new image samples, greatly improving the detection efficiency, accuracy and adaptability of dust pollution evaluation.

[0151] The units involved in the embodiments described in this disclosure may be implemented by software or hardware.

[0152] It should be understood that various parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0153] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A dust pollution evaluation method based on deep residual network, characterized in that: include: Step 1: Collect multiple sets of dust images after uniform diffusion of dust of different masses and arrange them in order to form an initial image set; Step 2: Preprocess each group of images in the initial image set to obtain a binary image with each pixel value being 1 or 0. Pixel points with pixel values ​​of 1 and 0 represent dust particles and gaps between particles or other background elements, respectively, to obtain a binary distribution image set. Step 3, respectively calculating the grayscale mean of each group of images in the binary distribution image set to obtain a grayscale mean result set; Step 4: Screen the grayscale mean result set according to statistical standards and remove abnormal data to obtain the grayscale mean determination set; The step 4 specifically includes: The Chauvenet criterion is used to determine whether the absolute value of the residual of the data in the gray mean result set satisfies the discriminant. The abnormal data are deleted from the gray mean result set, so that the gray mean retained set is all valid data and is used as the gray mean determination set. The discriminant of the Chauvenet criterion is: Among them, x i is the calculated grayscale mean data, Z c is the coefficient that varies according to the number of trials n, and are the arithmetic mean and standard deviation estimates of the grayscale mean of images under the same quality conditions; Step 5: Perform arithmetic mean calculation on the grayscale mean determination set to obtain the grayscale mean arithmetic mean set, and perform linear fitting on the grayscale mean set with the dust mass to obtain a linear function relationship between the grayscale mean and the dust mass; The step 5 specifically includes: In the grayscale mean determination set, the grayscale means of the same order between groups are selected, and the arithmetic mean of the grayscale means corresponding to the same test mass is calculated. The one-to-one correspondence between dust mass and arithmetic mean is expressed as the following function: in, Represents the arithmetic mean of the grayscale mean of group order n, m n Indicates the mass of the dust test with group order n; Then, a linear fitting was performed on the relationship between the grayscale mean arithmetic mean and the dust mass to obtain the linear function relationship between the grayscale mean and the dust mass: Where a and b represent the slope and intercept of the linear relationship respectively; Step 6: Divide the measured dust mass interval into T dust mass grade intervals with equal spacing. Determine the pollution levels corresponding to the dust mass grade intervals in ascending order of quality. Then, calculate the grayscale mean grade intervals using the linear function relationship between the grayscale mean and the dust mass and correspond them to the pollution levels. Obtain T groups of grayscale mean grade partition sets. Obtain the initial image grade partition set using the correspondence between the grayscale mean and the initial image set. Step 7: Expand the image level classification set to obtain an image level expansion set; Step 8: Input the expanded image grade set as training data into a preset image classification convolutional neural network model, and adjust the parameters of the model to obtain a dust pollution grade evaluation model to rate the image to be detected; Before step 8, the method further includes: Construct a convolution layer with a convolution kernel size of fc×fc×Cc; Construct an activation layer after the convolutional layer using ReLU as the activation function; Construct a pooling layer with a pooling window size of fp×fp, including an average pooling layer and a maximum pooling layer; After the pooling layer, an activation layer is constructed using ReLU as the activation function; Three types of residual modules are constructed through convolutional layers and pooling layers, each containing x convolutional layers and y pooling layers; Construct a softmax layer; Construct two fully connected layers to form a preset image classification convolutional neural network model.

2. The method according to claim 1, characterized in that ,The preprocessing operations include obtaining image gradient, gray scaling, sharpening, binarization and opening ,operation denoising.

3. The method according to claim 1, characterized in that The expansion processing includes at least one of flipping, scaling, rotating, graying, sharpening or adjusting brightness.

4. The method according to claim 1, characterized in that , the calculation formula of the grayscale mean is: in, is the grayscale mean, Z1 is the number of pixels with grayscale value 1 in the binary image, and M and N are the height and width of the binary image input size respectively.

5. The method according to claim 4, characterized in that , the step 6 specifically includes: Calculate the mass value that divides the measured dust mass interval into equidistant intervals, divide the measured dust mass interval into T dust mass grade intervals, which correspond to pollution levels 1 to T from left to right, and obtain the coordinates of the dividing points through the linear relationship between the grayscale mean and the dust mass. Use the vertical coordinate as the grayscale mean that divides the grayscale mean grade interval, and obtain T grayscale mean grade intervals, which correspond to pollution levels 1 to T from left to right.

6. The method according to claim 5, characterized in that , the step 8 specifically includes: The image level expansion set is input into the preset image classification convolutional neural network model. The weight parameters W and bias parameters b of all matrix operations in the neural network are updated using the loss function L of the internal network layer and the gradient descent optimizer. After repeated iterations, the dust pollution level evaluation model is obtained. The image to be detected is input into the dust pollution level evaluation model to obtain the pollution rating result.

7. The method according to claim 6, characterized in that , the expression of the loss function is Where N is the total number of image samples, L ij is the loss function of a single image sample, y it is the screening function for correct prediction of image samples, p(i, j, t) is the pollution level prediction probability of the image sample, z(i, j, t) represents the pollution level prediction weight of the image sample obtained by neural network operation, i represents the pollution level of the level expansion set where the image sample is located, j represents the sequence number of the image sample in the level expansion set, t represents the pollution level predicted for the sample, and T is the number of pollution level divisions.

Images