CN106709903B

CN106709903B - PM2.5 concentration prediction method based on image quality

Info

Publication number: CN106709903B
Application number: CN201611031140.0A
Authority: CN
Inventors: 陈强; 杨本芊; 徐琳
Original assignee: Nanjing University of Science and Technology
Current assignee: Nanjing University of Science and Technology
Priority date: 2016-11-22
Filing date: 2016-11-22
Publication date: 2020-06-19
Anticipated expiration: 2036-11-22
Also published as: CN106709903A

Abstract

The invention discloses a PM2.5 concentration prediction method based on image quality, and belongs to the technical field of image processing. The method comprises the steps of firstly carrying out image registration on a collected data set (screening images shot in rainy days), and picking image blocks meeting the dark channel principle to serve as a final training set. And then restoring the transmissivity graph of the training set, extracting features of the transmissivity image by using a sliding window method, standardizing the image features, and eliminating the influence of relative humidity on the image features. And then analyzing the relation between the extracted features of the training set and the real PM2.5 concentration by using a robustness regression analysis method, and further obtaining a PM2.5 concentration prediction model. And finally, the PM2.5 concentration prediction is completed. Experimental results show that the PM2.5 concentration predicted by the algorithm provided by the invention is superior to that of the existing algorithm, the influence of the relative humidity meteorological conditions on atmospheric imaging can be overcome, and the method has important significance for monitoring the PM2.5 concentration in daily life.

Description

PM2.5 concentration prediction method based on image quality

Technical Field

The invention relates to a PM2.5 concentration prediction method, in particular to a PM2.5 concentration prediction method based on image quality.

Background

PM2.5 is a general term for all particulate matters with the diameter less than or equal to 2.5 μm suspended in the air, is an important index for measuring the air quality, and has great influence on the human health. Currently, the detection of PM2.5 concentration in air is performed by large equipment, which is expensive and requires regular maintenance. The observation shows that the quality of the images shot under different air quality conditions is obviously different. Therefore, a model construction method for correlation between image quality and PM2.5 concentration is designed. The existing PM2.5 concentration prediction method based on image processing mainly extracts some features capable of reflecting image quality. In recent years, the following two image processing-based PM2.5 concentration prediction methods have mainly appeared:

(1) method based on visual features of images. The method extracts the gradient and color characteristics of the image. The PM2.5 concentration is estimated by the sky color difference of the taken image, and the method is influenced by weather, such as: the yin sky is heavy, increasing the error of estimation. The effect of relative humidity on image quality is not taken into account by the gradient characteristics of the image.

(2) Method based on physical features of images. The method utilizes an atmospheric imaging physical model, recovers a transmittance graph of an image by adopting a dark primary color prior estimation method, extracts a characteristic matrix from the transmittance graph by using a sliding window strategy, and establishes a relation model between the characteristic matrix and the real PM2.5 concentration by utilizing a robustness regression analysis method. But this method does not take into account the effect of relative humidity on atmospheric imaging. The relative humidity affects the extinction capability of the particulate matter PM2.5 in the air, and the larger the relative humidity is, the more water the PM2.5 in the air absorbs, the stronger the scattering capability to the atmospheric light is, and the more blurred the image is.

Therefore, the existing PM2.5 prediction method based on image processing does not consider the influence of relative humidity on atmospheric imaging, resulting in low accuracy of PM2.5 concentration estimation.

Disclosure of Invention

The invention aims to provide a novel method for constructing an image quality and PM2.5 concentration correlation model.

The technical solution for realizing the purpose of the invention is as follows: a novel method for constructing an image quality and PM2.5 concentration correlation model comprises the following steps:

step 1, collecting natural images at fixed points and at fixed time, and preprocessing the images;

step 2, extracting a transmissivity graph of the collected image by adopting a defogging algorithm based on a dark channel prior theory;

step 3, extracting a characteristic matrix from the transmittance image obtained in the step 2 by using a sliding window method;

step 4, carrying out standardization processing on the characteristic matrix obtained in the step 3, and eliminating the influence of relative humidity on the characteristic matrix;

step 5, modeling by using a robust regression analysis method to obtain a PM2.5 concentration prediction model;

and 6, predicting the PM2.5 concentration by using the model obtained in the step 5.

Compared with the prior art, the invention has the following remarkable advantages: 1) the method considers and overcomes the influence of the relative humidity on the atmospheric imaging, and improves the prediction accuracy of the PM2.5 estimation. 2) The method only deducts blocks which accord with the dark channel principle in the image to establish a prediction model; useless blocks in the image are simply and effectively removed, the modeling time is greatly shortened, and meanwhile, the memory space required by modeling is also reduced.

The present invention is described in further detail below with reference to the attached drawing figures.

Drawings

FIG. 1 is a flow chart of a new method of estimating PM2.5 concentration based on image processing according to the present invention.

FIG. 2 is a flow chart of the effect of normalized relative humidity on image transmittance characteristics.

Fig. 3 is a natural image collected with a smartphone.

Fig. 4 is a useful image block scratched from an original image.

Fig. 5 is a natural image transmittance graph.

FIG. 6 is a diagram of empirical formula f (RH) of the scattering moisture absorption growth factor of particles in air.

Fig. 7 is a zone prioritization chart.

FIG. 8 is a training model obtained by the method of the present invention.

FIG. 9 is a training model derived from a method based on physical features of an image.

FIG. 10 is a plot of estimated PM2.5 concentration versus actual PM concentration for the method of the present invention.

FIG. 11 is a comparison of the estimated PM2.5 concentration and the true PM based on the method of physical image characteristics.

Detailed Description

With reference to fig. 1, the new method for estimating the PM2.5 concentration based on image processing of the present invention comprises the following steps:

step 1, collecting natural images at fixed points and at fixed time, and preprocessing the images; the method specifically comprises the following steps:

step 1-1, screening out images shot in a data set in rainy days;

step 1-2, carrying out image registration on the residual data set, selecting one image as a reference image to carry out registration on the other images, wherein a Generalized Dual Bootstrap-ICP algorithm is used during the registration, and a transformation model selects Similarity;

and 1-3, scratching a certain block of an image in the data set as a final training set, and removing useless information of the image, wherein the image block contains a scene meeting the dark channel principle.

Step 2, extracting a transmissivity graph of the collected image by adopting a defogging algorithm based on a dark channel prior theory; the method specifically comprises the following steps:

step 2-1: and respectively carrying out minimum value filtering on three channels of the image R, G, B in the training set, wherein the window size p is as follows:

1).p＝m*m

2).m＝floor(max([3,w*kenlRatio,h*kenlRatio]))

wherein m is the window diameter, w is the width of the image, h is the height of the image, and kenlRatio is a proportion with the value between 0.01 and 0.05; after minimum filtering is carried out on the three channels, the minimum brightness value of a pixel in the three channels is selected as the brightness value of a pixel point corresponding to the dark channel map, and therefore the dark channel map is restored;

step 2-2: the atmospheric illumination intensity a of each image was found: firstly, taking the first 0.1% of pixels from a dark channel according to the brightness; secondly, in the positions, the corresponding value of the point with the highest brightness is searched in the original foggy image I and is used as an A value;

step 2-3: constructing an atmospheric imaging physical model, namely I (x) ═ J (x) t (x) + A (1-t (x)), wherein x is a pixel point coordinate, I is an observed foggy image, J is a clear fogless image, A is global atmospheric illumination intensity, and t is used for describing the part of light which is not scattered in the process of transmitting the light to imaging equipment through a medium and the transmittance;

step 2-4: the dark channel prior theoretical model is constructed as follows:

wherein x is the pixel point coordinate, c represents any channel, y is the domain pixel point coordinate of pixel point x, J is the clear fog-free image, J^duck(x) Is a dark channel map, and the model shows that in most of the local areas other than the sky, some pixels always have very low values, which tend to be 0;

step 2-5: deriving a formula according to the formula

Wherein x is a pixel point coordinate, c represents an arbitrary channel, y is a field pixel point coordinate of the pixel point x, I is an observed foggy image, and t (x) is a transmissivity image to be solved; and recovering the coarse transmittance graph through the formula, refining the coarse transmittance graph by using an instructive filter, wherein the refined transmittance graph is the required transmittance graph.

The transmittance image generally conforms to the law of near-far and near-far, because the particles in the air are uniformly distributed, the distance of the light emitted by an object close to the image in the air is short, more light enters the camera, and the corresponding transmittance value is large. Conversely, the transmittance value of the object at a far position is small.

Step 3, extracting a characteristic matrix from the transmittance image obtained in the step 2 by using a sliding window method; the method specifically comprises the following steps:

step 3-1, setting the size of the sliding window to be

h and w are the height and width of the image, ws is the window size, and the step of movement is set to

step is the step size of the window sliding;

and 3-2, gradually moving the sliding window along the transverse and longitudinal directions of the transmissivity image, and calculating the logarithm of the average value of the window brightness to be used as the characteristic value of the window, so as to obtain a characteristic matrix, wherein one image corresponds to one characteristic matrix.

Step 4, carrying out standardization processing on the characteristic matrix obtained in the step 3, and eliminating the influence of relative humidity on the characteristic matrix; the method specifically comprises the following steps:

step 4-1, determining the particle scattering moisture absorption growth factor empirical formula f (RH) ═ 1+ a (RH/100) in the air of different areas^bThe values of the two parameters a and b, wherein RH is relative humidity, and the values of the parameters a and b are shown in Table 1:

TABLE 1

Type of particulate matter in air	a	b
			City type	2.06	3.60
Ocean/city hybrid	3.26	3.85
			Ocean type	4.92	5.04

And 4-2, normalizing the feature matrix by using the determined f (RH), namely dividing each feature value in the feature matrix by f (RH) to eliminate the influence of the relative humidity on the feature value.

Step 5, modeling by using a robust regression analysis method to obtain a PM2.5 concentration prediction model; the method specifically comprises the following steps:

step 5-1, analyzing the relation between the characteristic value of each window of the training set and the real PM2.5 concentration by using a robustness regression analysis method, and calculating the correlation between the characteristic value and the real PM2.5 concentration;

and 5-2, selecting a window with the highest correlation with the real PM2.5 concentration as an optimal window, and taking a relation model corresponding to the window as a final training model to estimate the PM2.5 concentration, wherein the model comprises four parts in total, namely the relation model, the optimal window coordinate, the window size ws and the moving step size step.

And 6, predicting the PM2.5 concentration by using the model obtained in the step 5. The prediction model can only be used for predicting PM2.5 concentration corresponding to images shot in the local time collected by the training set, and specifically comprises the following steps:

6-1, extracting a feature matrix of the image to be detected according to the steps;

and 6-2, taking the corresponding characteristic value of the optimal window of the image to be measured as input and transmitting the input to a prediction model, and returning a predicted value to the prediction model, namely the PM2.5 concentration of the image to be measured at the local time of shooting.

The method considers and overcomes the influence of the relative humidity on the atmospheric imaging, and improves the prediction accuracy of the PM2.5 estimation. The method only extracts blocks which accord with the dark channel principle in the image to establish a prediction model; useless blocks in the image are simply and effectively removed, the modeling time is greatly shortened, and meanwhile, the memory space required by modeling is also reduced.

The present invention will be described in further detail with reference to the following examples:

the system takes a natural image shot by a smart phone as input, and adopts an image processing means to predict the image shot by the input image to obtain the local PM2.5 concentration. In order to predict the density of PM2.5 at the time of the input image shooting, firstly, images are acquired by a smart phone at the input image shooting time every day, and the number of the images is more than 15. Then, the method dynamically constructs a prediction model, and finally, the PM2.5 concentration of the input image at the local time can be predicted.

The flow of this embodiment is shown in fig. 1, the size of the fixed-point timing natural color image collected by the smartphone imaging device is 4160 × 3120, the total number of images is 20, and the captured image is shown in fig. 3. Images taken in rainy days are first screened, and then the remaining 15 images are subjected to image registration. The image registration is performed by using the Generalized Dual Bootstrap-ICP algorithm, and the Similarity is selected by the transformation model during the registration. Following the matting of the tiles, FIG. 4 shows the tiles actually used in this example, which are 400 × 300 in size.

After the 15 images are registered and preprocessed in a matting way, a dark channel prior theory-based defogging algorithm is adopted to recover the transmissivity of the 15 images, and as shown in fig. 5, the method specifically comprises the following steps:

the first step is as follows: minimum filtering is performed on three channels of the 15 images R, G, B, and the window size is 20 × 20. And selecting the minimum brightness value of the pixel in the three channels as the brightness value of the pixel point corresponding to the dark channel image.

The second step is that: the atmospheric illumination intensity a of each image was found: firstly, taking the first 0.1% of pixels from a dark channel according to the brightness; second, in these positions, the value of the corresponding point with the highest luminance is found in the original foggy image I as the a value.

The third step: according to the formula

And restoring the coarse transmittance graph, and then refining the coarse transmittance graph by using an instructive filter, wherein the refined transmittance graph is the required transmittance graph.

And after the transmissivity graph is restored, extracting a characteristic matrix from the transmissivity image by using a sliding window method. The method comprises the following specific steps:

the first step is as follows: the sliding window size is set to 15 x 15 in this example, and the window movement step size is 2.

The second step is that: and gradually moving the sliding window along the transverse and longitudinal directions of the transmissivity image, and calculating the logarithm of the average value of the window brightness to be used as the characteristic value of the window, thereby obtaining the characteristic matrix.

And extracting the feature matrix of the 15 images, and carrying out standardization processing on the feature matrix to eliminate the influence of relative humidity on the feature matrix. Fig. 6 shows a schematic diagram of an empirical formula f (rh) of a scattering moisture absorption growth factor of particles in air. Overall, f (RH) increases with increasing RH, with values for RH < 40% and f (RH) approaching 1, indicating that there is no significant increase in particle size of the airborne particles at lower ambient relative humidity. In addition, f (rh) in different regions has a distinct difference under the same humidity conditions. In this example, a is 1.24 and b is 6.27, the resulting feature matrix is normalized according to the steps given in fig. 2 to eliminate the effect of relative humidity.

And then modeling by using a robust regression analysis method to obtain a PM2.5 concentration prediction model. In the example, 15 images are collected, the PM2.5 concentration of each image at the local time is estimated by adopting the remaining method, namely 14 images are used as a training set for modeling each time, and the remaining image is used as a test set for testing for 15 times. The method comprises the following specific steps:

the first step is as follows: a robust regression analysis method (matlab self-contained function robustfit function) is used for analyzing the relationship between the characteristic value (each characteristic value in the characteristic matrix) of each window of the 14 images and the real PM2.5 concentration to construct a prediction model, and the correlation between the characteristic value and the real PM2.5 concentration is calculated to obtain a region priority sequence diagram, as shown in fig. 7, each window corresponds to one model correlation, and the darker the color represents the higher the correlation. The window with the deepest color is marked as an optimal window, the coordinate of the optimal window is 23, and the coordinate of the optimal window is 90.

The second step is that: the relationship model corresponding to the optimal window x-23 and y-90 (as shown in fig. 8) is used as the final training model to perform PM2.5 concentration estimation.

And finally, performing PM2.5 concentration prediction on the test set image by using the obtained prediction model, specifically, substituting the characteristic value corresponding to the optimal window (x is 23, y is 90) of the test set into the prediction model, and returning a PM2.5 concentration prediction value by the prediction model.

Fig. 9 shows a PM2.5 concentration prediction model obtained by a conventional method for extracting physical features of an image. FIG. 10 is a comparison of the estimated PM2.5 concentration value obtained in the present example with the true value. FIG. 11 is a comparison of the estimated PM2.5 concentration and the true PM based on the method of physical image characteristics. Table 2 compares the process of the invention with known processes.

TABLE 2

	Before standardization	After standardization	Improvements in or relating to
				R	0.8741	0.9056	0.0315
MAE(ug/m3)	7.0219	5.7105	1.3114

As can be seen from fig. 10, fig. 11, and table 2: the PM2.5 concentration prediction model obtained by the method can accurately estimate the PM2.5 concentration of the smart phone shot image at the current time and the local place, overcomes the influence of relative humidity on atmospheric imaging, is superior to the existing method in the correlation R and the absolute average error MAE between the PM2.5 concentration estimated value and the true value, and only blocks which accord with the dark channel principle in the image are deducted to establish the prediction model. Useless blocks in the image are simply and effectively removed, and the modeling time is greatly shortened. The shortening of the time and the improvement of the estimation precision provide convenience for the prediction and monitoring of the PM2.5 concentration in daily life.

Claims

1. A PM2.5 concentration prediction method based on image quality is characterized by comprising the following steps:

step 1-1, screening out images shot in a data set in rainy days;

step 1-3, a certain block of an image in a data set is scratched to serve as a final training set, useless information of the image is removed, and the image block contains a scene meeting a dark channel principle;

and 2, restoring a transmittance map of the acquired image by adopting a defogging algorithm based on a dark channel prior theory, which specifically comprises the following steps:

1).p＝m*m

2).m＝floor(max([3,w*kenlRatio,h*kenlRatio]))

step 2-4: the dark channel prior theoretical model is constructed as follows:

wherein x is the pixel point coordinate, c represents any channel, y is the domain pixel point coordinate of pixel point x, J is the clear fog-free image, J^dark(x) Is a dark channel map, the model shows that in most non-sky local areas, some pixels will always tend to 0;

step 2-5: deriving a formula according to the formula

Wherein x is a pixel point coordinate, c represents an arbitrary channel, y is a field pixel point coordinate of the pixel point x, I is an observed foggy image, and t (x) is a transmissivity image to be solved; recovering the coarse transmittance graph through the formula, refining the coarse transmittance graph by using an instructive filter, wherein the refined transmittance graph is the required transmittance graph;

2. The image quality-based PM2.5 concentration prediction method according to claim 1, wherein the step 3 of extracting the feature matrix for the transmittance image by using the sliding window method specifically comprises:

step 3-1, setting the size of the sliding window to be

step is the step size of the window sliding;

3. The image quality-based PM2.5 concentration prediction method according to claim 1, wherein the step 4 normalizes the feature matrix, specifically:

TABLE 1

4. The image quality-based PM2.5 concentration prediction method according to claim 1, wherein the step 5 of modeling with a robust regression analysis method to obtain the PM2.5 concentration prediction model specifically comprises:

5. The image quality-based PM2.5 concentration prediction method according to claim 1, wherein the step 6 uses the obtained model to predict the PM2.5 concentration, and the prediction model can only be used to predict the PM2.5 concentration corresponding to the image captured at that time in the training collection, specifically: