An intelligent image perception method for flue gas in flare venting

Abstract

The present invention discloses a method for intelligent image perception of flue gas for venting torches, wherein the implementation of the method includes: collecting venting torch images, performing black smoke presence detection, extracting image smoke regions using a saliency method, removing image noise using a Gaussian filter, improving image quality using a super-resolution method, and calculating the blackness level of the flue gas to analyze the quality of the flue gas. The present invention concentrates the target area around the flue gas and flame, reduces the influence of other objects around the venting torch, and enhances the anti-interference capability; performs image denoising by using a Gaussian filter, reduces the grayscale of image noise points, and thus reduces the amount of noise in the image; maps the flue gas image to a high-quality image by using a super-resolution method, which facilitates the analysis of the quality of the flue gas in the image and improves the accuracy of the determination result.

Description

Technical Field

[0001] The present invention relates to the field of image processing methods, and more specifically, to a flue gas intelligent image perception method for a flare. Background Art

[0002] As the country pays more and more attention to environmental protection, the requirements for enterprises to meet environmental protection standards are becoming more and more stringent. The flue gas generated by the flare systems of petrochemical plants, refineries and other enterprises has gradually been taken seriously. Most flue gas sensors are difficult to deploy on elevated flares, and are often easily damaged by high temperatures, and the cost is very high. Traditional detection methods are affected by factors such as sensors and cannot meet control requirements well. However, in the past few years, venting flare combustion systems based on flue gas detection have become increasingly popular, and image perception methods for flue gas detection have been widely studied.

[0003] Turgay et al. proposed a fire smoke detection model that combines color information and motion analysis. This method mainly extracts samples from different types of images or video sequences for statistical analysis and extracts relevant color models. The basic assumption of this fire smoke detection model is that when a fire occurs, the color of the smoke generated is predicted to change from white-blue to white, and as the fire continues, the color of the smoke will change close to the black area. The detection method includes low saturation detection in HSV space, some simple motion detection and threshold detection between RGB channels. However, due to the interference of the color of clouds and their simple motion in the background environment, this detection technology has a high false alarm rate.

[0004] Yu et al. proposed a smoke detection method based on wavelet features and neural networks. This method mainly detects smoke by analyzing the texture information of smoke and monitors fire in real time. This method mainly includes three main steps: ① Divide the acquired video frame into multiple blocks; ② Use adaptive Gaussian mixture model and background subtraction to determine the foreground pixels of the block, and select blocks with foreground pixels greater than a certain threshold as candidate blocks; ③ Perform GLCM feature extraction on all candidate blocks, generate features by calculating the contrast, uniformity and energy of the relevant candidate blocks, and use the extracted features for BP neural network classification. Although this method has achieved good results in the experiment, it is difficult to obtain the same experimental results in smoke texture sets of different sizes and different noise levels. In addition, some steps still need to be optimized to achieve real-time detection effects.

[0005] It can be seen that the existing smoke quality perception related technologies are affected by image noise and surrounding interference, and have a high error rate in smoke quality analysis. Summary of the invention

[0006] In order to solve the problems that the above-mentioned flue gas quality detection process is susceptible to noise interference and the error rate of flue gas quality judgment is high, the main purpose of the present invention is to provide a flue gas intelligent image perception method for venting torches, so as to solve the problem that flue gas monitoring based on image processing technology is susceptible to interference under environmental factors such as noise, and to improve the accuracy of the final judgment of flue gas quality.

[0007] To achieve the above object, the present invention provides a flue gas intelligent image perception method for flare venting, comprising the following steps:

[0008] Step 1: Image acquisition: Use a high-definition camera to capture the burning conditions of the flare to complete the image acquisition;

[0009] Step 2: Detection of the presence of black smoke: Use the IDCNN network to identify whether there is black smoke in the flare image. If black smoke is detected, further determination of the quality of the black smoke is performed and the process proceeds to step 3. If no black smoke is detected, the following process is terminated.

[0010] Step 3: Extract the smoke area of ​​the image. Use the image salient area extraction method on the collected image to focus the research target on the area around the smoke.

[0011] Step 4: Noise removal: Use Gaussian filtering method to remove noise points in the image after extracting the smoke area;

[0012] Step 5: Improve the image quality by using the super-resolution method to reconstruct and map the denoised image in step 4 to improve the image quality;

[0013] Step 6: Determine the smoke quality. Perform smoke grayscale calculation and analysis on the image in step 5 and output its blackness level. The blackness level represents the smoke quality.

[0014] Preferably, the collected image may be a picture directly taken by a high-definition camera, or may be a frame of video image captured from a video of monitoring the flare venting.

[0015] Preferably, the IDCNN network described in step 2 is an end-to-end deep dual-channel neural network, which is formed by fusing a deep subnet ISBNN that is good at extracting detailed features of smoke and a deep subnet ISCNN that is good at extracting basic information of smoke. The image collected in step 1 is input into the IDCNN network to determine whether there is obvious black smoke generated in the collected image;

[0016] Among them, the ISBNN network and the ISCNN network are improved from the two sub-networks of DCNN (Deep Dual-Channel Neural Network), the SBNN network and the SCNN network. The fusion method of the ISBNN network and the ISCNN network is the same as the fusion method of the SBNN network and the SCNN network into the DCNN network; the network structure of the SBNN is: convolution layer L1, convolution layer L2, maximum pooling layer L3, normalization and convolution layer L4, normalization and convolution layer L5, maximum pooling layer L6, normalization and convolution layer L7, normalization and convolution layer L8, maximum pooling layer L9, fully connected layer L10, fully connected layer L11, output layer L12; the network structure of the SCNN is: convolution layer M1, convolution layer M2, normalization and convolution layer M3, normalization and convolution layer M4, normalization and convolution layer M5, cascade and convolution layer M6, maximum pooling layer M7, normalization and convolution layer Normalization and convolution layer M8, normalization and convolution layer M9, maximum pooling layer M10, normalization and convolution layer M11, normalization and convolution layer M12, convolution layer M13, maximum pooling layer M14; in the SBNN network structure, the ordinary convolutions in L2, L5 and L7 are replaced by depthwise separable convolution to construct an ISBNN sub-network, and the ordinary convolutions in M3, M4, M5, M6, M9 and M12 are replaced by depthwise separable convolution to construct an ISCNN sub-network; the parameters of the introduced depthwise separable convolution layer are consistent with the corresponding parameters of the replaced ordinary convolution layer; the ISBNN and ISCNN networks are fused to generate the IDCNN network according to the fusion method of the DCNN network; the specific network layers before and after the improvement can be seen Figure 1 , Figure 2 ;

[0017] In the IDCNN network training, ISBNN and ISCNN are trained independently;

[0018] During the ISBNN training process, the glorot uniform method was introduced to initialize the network weights, and the stochastic gradient descent method was used to train the ISBNN by setting the momentum coefficient to 0.9, the initial learning rate to 0.01, and the learning rate decay coefficient to 0.0001;

[0019] The training of ISCNN performs the same process as above;

[0020] Finally, the two are combined and the overall parameters of IDCNN are fine-tuned to search for the optimal parameters.

[0021] Preferably, a method for extracting a salient region of an image comprises the following steps:

[0022] (1) Image preprocessing: The original captured image is divided into 200 superpixel blocks and Lab color features are extracted from them;

[0023] (2) Saliency value calculation: Calculate the spatial distance between each superpixel block and other superpixel blocks, and use these spatial distances to calculate the saliency value of each superpixel block, thereby obtaining the basic saliency map of the collected image;

[0024] (3) Obtaining the final saliency map: Based on the similarity metric between superpixel blocks, the saliency values ​​obtained in step (2) are weighted optimized using the sequence synthesis method to obtain the final saliency map of the input image.

[0025] (4) Extraction of salient regions: The final salient map of the acquired image is subjected to image adaptive threshold segmentation using the Otsu algorithm to extract the salient regions of the acquired image.

[0026] Preferably, the significant region image extracted in step 3 is first enlarged to the original image size, and a Gaussian filtering method is used to remove obvious noise points in the image.

[0027] Preferably, in step 5, a SRCNN super-resolution network is used on the image to learn an end-to-end mapping from a low-resolution image to a high-resolution image;

[0028] The sizes of convolution kernels used in the three convolutional layers of the SRCNN network are 9x9, 1x1, and 5x5, and the number of output features of each layer is 64, 32, and 3 respectively; the Timofte dataset (containing 91 images) and the ImageNet large dataset are used for training; the mean square error (MSE) is used as the loss function;

[0029] The main process of the SRCNN super-resolution network taking a low-resolution image as input and finally outputting a high-resolution image is as follows:

[0030] (1) Extract features from the original low-resolution image and represent it as a high-dimensional feature vector;

[0031] (2) Nonlinearly map the high-dimensional feature vector to another set of high-dimensional feature vectors;

[0032] (3) Reconstruct high-resolution images.

[0033] Preferably, the smoke blackness calculation step in step 6 is:

[0034] (1) For the smoke area corresponding to the image reconstructed in step 5, find the darkest point in the smoke area, that is, find the point in the smoke area with the shortest color distance to black RGB (0,0,0), where the RGB color distance formula is:

[0035]

[0036] Among them, Y1 corresponds to the color value of each point in the smoke area of ​​the image, Y2 corresponds to the color value of RGB (0,0,0), and Y 1,R , Y 1,G , Y 1,B Represents the RGB three channel values ​​of Y1, Y 2,R , Y 2,G , Y 2,B Same as above;

[0037] After finding the point, convert the RGB value of the point into a grayscale value. The relevant conversion formula is:

[0038] Gray=R*0.299+G*0.587+B*0.114,

[0039] Among them, R, G, B represent the values ​​of the RGB channels of the pixel respectively, and Gray represents the grayscale value of the pixel;

[0040] (2) comparing the gray value obtained in step (1) with the gray value corresponding to the Ringelmann blackness level, and setting the Ringelmann blackness level corresponding to the minimum difference between the gray value and the Ringelmann blackness level as the blackness level of the smoke area;

[0041] When observing smoke under cloudy conditions, due to the dark sky background, according to past experience, a slightly lower level should be read when reading the blackness level, such as subtracting 0.25 or 0.5 from the read level, and finally outputting the smoke area blackness level;

[0042] The Ringelmann blackness level is a flue gas quality measurement indicator. The smoke quality is evaluated by further combining the smoke area blackness level calculated above with the smoke blackness emission limit.

[0043] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0044] (1) The present invention can focus the target area around smoke and flames by using an image saliency algorithm, thereby reducing the influence of other objects around the flare and enhancing the anti-interference capability;

[0045] (2) By using Gaussian filtering to eliminate Gaussian noise that exists in most images, this method can reduce the grayscale of noise points, thereby reducing the amount of noise in the image;

[0046] (3) By using the super-resolution method to map the smoke image into a high-quality image, it is easier to analyze the smoke quality in the image and improve the accuracy of the judgment result. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1It is the various network layers in SBNN and SCNN networks;

[0048] Figure 2 It is the various network layers in ISBNN and ISCNN networks;

[0049] Figure 3 It is a framework diagram of the smoke intelligent image perception method for flare venting provided by the present invention; DETAILED DESCRIPTION

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are further described below in conjunction with the accompanying drawings.

[0051] 1. Image Acquisition

[0052] The present invention perceives the quality of flare smoke based on image processing, so relevant images need to be collected before image processing and analysis. First, start the high-definition camera and the computer device connected to it, collect images in real time through the camera, and transmit the real-time collected images to the computer device. There are two specific collection methods:

[0053] (1) Use a high-definition camera to directly shoot the flare to obtain relevant images;

[0054] (2) Video monitoring of the flare is performed by capturing a certain frame of the video as the collected image. The capture method is equidistant capture, such as obtaining a frame of the image every one second.

[0055] 2. Detection of black smoke presence

[0056] Before analyzing the flue gas quality, it is necessary to determine whether the current venting torch emits obvious black smoke. If the venting torch does not generate black smoke, the current process ends; if obvious black smoke is detected, the flue gas quality analysis continues.

[0057] The present invention adopts the IDCNN network to detect the presence of black smoke. The IDCNN network is an end-to-end deep dual-channel neural network, which is formed by the fusion of the deep subnet ISBNN that is good at extracting the detailed features of smoke and the deep subnet ISCNN that is good at extracting the basic information of smoke. The specific implementation steps are as follows:

[0058] (1) IDCNN network training

[0059] In the IDCNN network training, ISBNN and ISCNN were trained independently;

[0060] In the ISBNN training process, the optimal network structure is first found through trial and error, the glorot uniform method is introduced to initialize the network weights, and the stochastic gradient descent method is used to train the ISBNN by setting the momentum coefficient to 0.9, the initial learning rate to 0.01, and the learning rate decay coefficient to 0.0001;

[0061] The training of ISCNN performs the same process as above;

[0062] Finally, the two are combined and the overall parameters of IDCNN are fine-tuned to search for the optimal parameters.

[0063] (2) Detection of black smoke presence

[0064] After the IDCNN network is trained, the collected images are input into the IDCNN network to determine whether there is obvious black smoke generated in the collected images.

[0065] 3. Extract the smoke area of ​​the image

[0066] After determining that there is obvious black smoke in the collected image, in order to focus the research target on the area around the smoke, the smoke area of ​​the flare is extracted from the collected image, which involves a method for extracting image salient areas, including the following steps:

[0067] (1) Image preprocessing: The original captured image is divided into 200 superpixel blocks and Lab color features are extracted from them;

[0068] (2) Saliency value calculation: Calculate the spatial distance between each superpixel block and other superpixel blocks, and use these spatial distances to calculate the saliency value of each superpixel block, thereby obtaining the basic saliency map of the collected image;

[0069] (3) Obtaining the final saliency map: Based on the similarity metric between superpixel blocks, the saliency values ​​obtained in step (2) are weighted optimized using the sequence synthesis method to obtain the final saliency map of the input image.

[0070] (4) Extraction of salient regions: The final salient map of the acquired image is subjected to image adaptive threshold segmentation using the Otsu algorithm to extract the salient regions of the acquired image.

[0071] 4. Noise Removal

[0072] For the image after smoke extraction, noise removal is performed. The specific steps are as follows:

[0073] (1) Image enlargement: The image of the salient area after smoke extraction is enlarged to the original image size using the bicubic interpolation method;

[0074] (2) Image denoising: Gaussian filtering is used to remove obvious noise points in the image. In the Gaussian denoising process, a Gaussian filter that conforms to the Gaussian function distribution is selected to perform weighted averaging on the image.

[0075] The size of the Gaussian filter template is selected as 5×5, the standard deviation σ is selected as 1.1, the center of the 5×5 template is used as the coordinate origin (0,0), and the corresponding integer coordinates are generated for the remaining positions of the template;

[0076] Substitute the coordinates of each position into the Gaussian function to obtain the corresponding coefficients of the template. The calculation formula for each element value in the Gaussian function template is as follows:

[0077]

[0078] Where (i, j) is the position coordinate of the template element, σ is the standard deviation, and since the window template size here is 5, the k value here is 2;

[0079] Finally, the center of the generated Gaussian template is aligned with the image matrix to be processed, and pixel value weighting processing is performed to obtain the result image after Gaussian filtering.

[0080] 5. Improve image quality

[0081] After noise removal, the image will be blurred to a certain extent, so the image quality needs to be further improved before smoke quality analysis. In the process of image quality improvement, the SRCNN super-resolution network is used on the image to learn the end-to-end mapping from low-resolution images to high-resolution images;

[0082] The sizes of convolution kernels used in the three convolutional layers of the SRCNN network are 9x9, 1x1, and 5x5, and the number of output features of each layer is 64, 32, and 3 respectively; the Timofte dataset (containing 91 images) and the ImageNet large dataset are used for training; the mean square error (MSE) is used as the loss function;

[0083] The main process of the SRCNN super-resolution network taking a low-resolution image as input and finally outputting a high-resolution image is as follows:

[0084] (1) Extract features from the original low-resolution image and represent it as a high-dimensional feature vector;

[0085] (2) Nonlinearly map the high-dimensional feature vector to another set of high-dimensional feature vectors;

[0086] (3) Reconstruct high-resolution images.

[0087] 6. Determine the quality of flue gas

[0088] Analyze the smoke quality of the image and calculate the blackness level of the smoke area.

[0089] The calculation steps of smoke blackness are as follows:

[0090] (1) For the image with improved quality after reconstruction, find the darkest point in the smoke area, that is, find the point in the smoke area with the smallest color distance to black RGB (0,0,0), where the RGB color distance formula is:

[0091]

[0092] Among them, Y1 corresponds to the color value of each point in the smoke area of ​​the image, Y2 corresponds to the color value of RGB (0,0,0), and Y 1,R , Y 1,G , Y 1,B Represents the RGB three channel values ​​of Y1, Y 2,R , Y 2,G , Y 2,B Same as above;

[0093] After finding the point, convert the RGB value of the point into a grayscale value. The relevant conversion formula is:

[0094] Gray=R * 0.299+G * 0.587+B * 0.114,

[0095] Among them, R, G, B represent the values ​​of the RGB channels of the pixel respectively, and Gray represents the grayscale value of the pixel;

[0096] (2) The Ringelmann blackness diagram has six levels, ranging from 0 to 5, wherein all white and all black correspond to Ringelmann blackness levels 0 and 5, and the remaining four levels are determined according to the percentage of the black stripes in the entire area. The black stripes account for 20% of the area, which is level 1; 40% for level 2; 60% for level 3; and 80% for level 4. According to existing data, the grayscale values ​​of levels 0 to 5 can roughly correspond to 255, 255, 204, 153, 102, 51, and 0. The grayscale value obtained in step (1) is compared with the grayscale value corresponding to the Ringelmann blackness level, and the Ringelmann blackness level corresponding to the minimum difference between the two is set as the blackness level of the smoke area.

[0097] When observing smoke under cloudy conditions, due to the dark sky background, according to past experience, a slightly lower level should be read when reading the blackness level, such as subtracting 0.25 or 0.5 from the read level, and finally outputting the smoke area blackness level;

[0098] The Ringelmann blackness level is a flue gas quality measurement indicator. The smoke quality is evaluated by combining the smoke area blackness level calculated above with the smoke blackness emission limit.

[0099] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A flue gas intelligent image perception method for flare venting, characterized in that: The steps include: Step 1: Image acquisition: Use a high-definition camera to capture the burning conditions of the flare to complete the image acquisition; Step 2: Detection of the presence of black smoke: Use the IDCNN network to identify whether there is black smoke in the flare image. If black smoke is detected, further determination of the quality of the black smoke is performed and step 3 is entered; If no black smoke is detected, the following process is terminated; Step 3: Extract the smoke area of ​​the image. Use the image salient area extraction method on the collected image to focus the research target on the area around the smoke. Step 4: Noise removal: Use Gaussian filtering method to remove noise points in the image after extracting the smoke area; Step 5: Improve the image quality by using the super-resolution method to reconstruct and map the denoised image in step 4 to improve the image quality; Step 6: Determine the smoke quality, perform smoke grayscale calculation and analysis on the image reconstructed in step 5, and output its blackness level, which represents the smoke quality; The IDCNN network described in step 2 is an end-to-end deep dual-channel neural network, which is formed by fusing a deep subnet ISBNN that is good at extracting detailed features of smoke and a deep subnet ISCNN that is good at extracting basic information of smoke. The image collected in step 1 is input into the IDCNN network to determine whether there is obvious black smoke generated in the collected image; Among them, the ISBNN network and the ISCNN network are improved from the two sub-networks SBNN network and SCNN network of DCNN (Deep Dual-Channel Neural Network). The fusion method of the ISBNN network and the ISCNN network is the same as the method of fusion of the SBNN network and the SCNN network into the DCNN network; the network structure of the SBNN is: convolution layer L1, convolution layer L2, maximum pooling layer L3, normalization and convolution layer L4, normalization and convolution layer L5, maximum pooling layer L6, normalization and convolution layer L7, normalization and convolution layer L8, maximum pooling layer L9, full connection layer L10, full connection layer L11, output layer L12, and the network structure of the SCNN is: convolution layer M1, convolution layer M2, normalization and convolution layer M3, normalization and convolution layer M4, normalization and convolution layer M5, cascade and convolution layer M6, maximum pooling layer Layer M7, normalization and convolution layer M8, normalization and convolution layer M9, maximum pooling layer M10, normalization and convolution layer M11, normalization and convolution layer M12, convolution layer M13, maximum pooling layer M14; the improvement is: in the SBNN network structure, the ordinary convolution in L2, L5 and L7 is replaced by the depthwise separable convolution to construct an ISBNN sub-network, and in the SCNN network structure, the ordinary convolution in M3, M4, M5, M6, M9 and M12 is replaced by the depthwise separable convolution to construct an ISCNN sub-network; the parameters of the introduced depthwise separable convolution layer are consistent with the parameters of the corresponding replaced ordinary convolution layer; the ISBNN and ISCNN networks are fused to generate the IDCNN network in the same way as the DCNN network; The image salient region extraction method described in step 3 comprises the following steps: (1) Image preprocessing: The original captured image is divided into 200 superpixel blocks and Lab color features are extracted from them; (2) Saliency value calculation: Calculate the spatial distance between each superpixel block and other superpixel blocks, and use these spatial distances to calculate the saliency value of each superpixel block, thereby obtaining the basic saliency map of the collected image; (3) Obtaining the final saliency map: Based on the similarity metric between superpixel blocks, the saliency values ​​obtained in step (2) are weighted optimized using the sequence synthesis method to obtain the final saliency map of the input image. (4) Extraction of salient regions: The final salient map of the acquired image is subjected to image adaptive threshold segmentation using the Otsu algorithm to extract the salient regions of the acquired image.

2. The method for intelligent image perception of flue gas for flare venting according to claim 1, characterized in that: The collected image can be a picture taken directly by a high-definition camera, or a frame of video image captured from a video monitoring the flare venting.

3. The method for intelligent image perception of flue gas for flare venting according to claim 1, characterized in that: In the IDCNN network training, ISBNN and ISCNN are trained independently; During the ISBNN training process, the glorot uniform method was introduced to initialize the network weights, and the stochastic gradient descent method was used to train the ISBNN by setting the momentum coefficient to 0.9, the initial learning rate to 0.01, and the learning rate decay coefficient to 0.0001; The training of ISCNN follows the same process as above; Finally, the two are combined and the overall parameters of IDCNN are fine-tuned to search for the optimal parameters.

4. The method for intelligent image perception of flue gas for flare venting according to claim 1, characterized in that: Step 4 specifically includes firstly enlarging the salient region image extracted in step 3 to the original image size, and then using Gaussian filtering to remove noise in the image.

5. The method for intelligent image perception of flue gas for flare venting according to claim 1, characterized in that: The super-resolution method in step 5 specifically adopts the SRCNN super-resolution network to learn the end-to-end mapping from low-resolution images to high-resolution images.

6. The method for intelligent image perception of flue gas for flare venting according to claim 5, characterized in that: The sizes of convolution kernels used in the three convolutional layers of the SRCNN network are 9x9, 1x1 and 5x5, and the number of output features of each layer is 64, 32 and 3 respectively; the Timofte dataset and the ImageNet large dataset are used for training; the mean square error MSE is used as the loss function; The process of SRCNN super-resolution network taking low-resolution images as input and finally outputting high-resolution images is as follows: (1) Extract features from the original low-resolution image and represent it as a high-dimensional feature vector; (2) Nonlinearly map the high-dimensional feature vector to another set of high-dimensional feature vectors; (3) Reconstruct high-resolution images.

7. The method for intelligent image perception of flue gas for flare venting according to claim 1, characterized in that: The steps for calculating the smoke blackness in step 6 are: (1) Find the darkest point in the reconstructed smoke area, that is, find the point in the smoke area with the shortest color distance to black RGB (0,0,0), where the RGB color distance formula is: Among them, Y1 corresponds to the color value of each point in the smoke area of ​​the image, Y2 corresponds to the color value of RGB (0,0,0), and Y 1,R , Y 1,G , Y 1,B Represents the RGB three channel values ​​of Y1, Y 2,R , Y 2,G , Y 2,B Represent the RGB three channel values ​​of Y2 respectively; After finding the point, convert the RGB value of the point into a grayscale value. The conversion formula is: Gray=R * 0.299+G * 0.587+B * 0.114, Among them, R, G, B represent the values ​​of the RGB channels of the point respectively, and Gray represents the grayscale value of the point; (2) Compare the grayscale value obtained in step (1) with the grayscale value corresponding to the Ringelmann blackness level, and the Ringelmann blackness level corresponding to the minimum difference between the two is set as the blackness level of the smoke area.

Images