A dust detection and concentration classification method based on prior information

Abstract

This invention provides a dust detection and concentration classification method based on prior information, including the construction of a dust detection and monitoring system, the creation of a dataset, and the optimization of the algorithm. Compared with other deep learning detection technologies, this invention obtains a grayscale image through the Fast-DMPHN image dehazing network. This grayscale image can provide the dust concentration, and it has strong prior information. This prior information can compensate for the spatial information lost by the U-net network, thereby improving the segmentation accuracy. Combining the dust concentration and the area of ​​the dust region, a comprehensive judgment is made and the dust concentration is classified. This invention can automatically detect dust in a region and determine the corresponding level. Based on the segmentation results and dust concentration, it can achieve precise dust removal. Compared with previous dust detection methods, it has the advantages of fast detection speed, high accuracy, and rapid and accurate dust removal, which improves efficiency and saves human capital, and has great value and significance.

Description

Technical Field

[0001] This invention relates to the field of port coal dust detection technology, and in particular to a dust detection and concentration classification method based on prior information. Background Technology

[0002] Currently, there are two main methods for dust control at docks: one is to keep the dust removal equipment running throughout the operation, and the other is to manually determine whether to turn the dust removal equipment on or off. The former has low economic efficiency, while the latter increases labor costs. This invention uses computer vision to automatically detect dust through a video monitoring system, replacing manual real-time detection and judgment of dust concentration levels. At the same time, it performs precise water spraying according to the segmented areas, and uses the obtained dust concentration as the control standard for the amount of water sprayed in the dusty areas, thereby saving labor costs and improving work efficiency. Summary of the Invention

[0003] This invention uses water spraying to suppress dust, and utilizes computer vision technology to automatically determine the location of dust and the amount of water sprayed, solving the problems of low efficiency and high cost in previous dust removal work.

[0004] The technical means employed in this invention are as follows:

[0005] A dust detection and concentration classification method based on prior information specifically includes the following steps:

[0006] Step 1: Set up a video monitoring system for dust from chute unloading, and determine the type of monitoring camera and its installation location;

[0007] Step 2: Obtain the dataset, which includes dust images of chute unloading under different working conditions, as well as dust-free images that are strictly paired with the dust images, and label the coal dust under different scenarios.

[0008] Step 3, send the dataset to The module obtains a grayscale image called the dust concentration map (i.e., the A channel). This dust concentration map has prior information. The prior information is used to guide the U-net segmentation network with an attention mechanism to supplement spatial information and improve segmentation performance. The original image in the dataset and the dust concentration map are stitched together to form an RGBA image. The resulting RGBA image is input into the U-net segmentation network module with an attention mechanism to finally obtain the dust segmentation map, i.e., the precise location of the dust.

[0009] Step 4 uses the dust concentration map (i.e., channel A) obtained from image dehazing as a representation of dust concentration.

[0010] Step 5: Based on the approximate mapping relationship between the dust concentration map and the actual dust concentration value, and the dust segmentation map, the dust concentration and concentration classification are realized.

[0011] A further improvement of the present invention is that step 3 specifically includes:

[0012] Step 3-1, the The module is an image dehazing module. The image is input into the Fast-DMPHN module to obtain a grayscale image (i.e., the A channel) as a dust concentration map. The dust concentration map, with prior information, guides the segmentation network to perform dust segmentation. It is then concatenated with the input image to obtain an RGBA four-channel image, which is then input into the U-net segmentation network with an attention mechanism.

[0013] Step 3-2: Add channel attention and spatial attention models to the U-net segmentation network. Specifically, after each downsampling step in the U-net segmentation network, CBAM (channel attention and spatial attention model) is added, resulting in a network with an added attention mechanism. Segmentation network;

[0014] Step 3-3: Input the obtained RGBA image into the U-net segmentation network with added attention to obtain the final segmentation image of the dust, i.e., the precise location of the dust.

[0015] Steps 3-4: The U-net segmentation network uses a loss function with boundary weights. The formula for calculating this loss function is as follows:

[0016]

[0017] in It is the softmax loss function. It is the label of each pixel. These are the weights of the pixels;

[0018]

[0019] in It is the weight that balances the proportions of the categories. This represents the distance between a pixel and its nearest dust particle. This represents the distance between a pixel and its second nearest dust particle. and It is a constant value.

[0020] A further improvement of the present invention is that the dust concentration estimation of the dust concentration map specifically includes:

[0021] pass The module-based defogging algorithm obtains an estimated value of dust concentration, and its calculation formula is as follows:

[0022]

[0023] in, J represents the pixel brightness value with spatial coordinate x in the image, J represents the restored image obtained after passing through the dehazing network, and I represents the network input image, which is an image with dust. Both I and J are grayscale images.

[0024] According to the atmospheric scattering model:

[0025]

[0026] get:

[0027]

[0028] in, Let A be the transmittance and A be atmospheric light. The pixel value is known, and It is inversely proportional to the dust concentration, which shows that It is directly proportional to the dust concentration, therefore... To represent dust concentration;

[0029] right Filtering is performed to obtain the dust concentration in the segmented region.

[0030] A further improvement of this invention is that it also includes using the PyTorch deep learning framework and the dust dataset created in step 2 to incorporate prior information and an attention mechanism in step 3. The segmentation network is trained to obtain the final .weights file for dust detection, and then OpenCV is used to implement the video detection algorithm.

[0031] A further improvement of the present invention is that the dust detection and concentration classification specifically includes:

[0032] The dust video monitoring system collects real-time video of the chute unloading operation, incorporating prior information and an attention mechanism. The segmentation network detects dust in each frame of the image, obtains the dust segmentation region, and calculates the concentration value of the dust in the segmented region.

[0033] The actual dust concentration was measured on-site using a laser dust concentration detector.

[0034] Repeat the above two steps in different scenarios to obtain an approximate mapping relationship between the actual dust concentration and the actual dust concentration in different scenarios.

[0035] A further improvement of the present invention is that, when detecting dust concentration in real time, the area of ​​the dust segmentation region is used as an important information for judging dust concentration, and the dust concentration level is judged by combining it with the concentration in the dust region. Water is sprayed on the dust region according to the segmentation region, and the above-mentioned dust concentration is used as the control standard for controlling the amount of water sprayed.

[0036] Due to the adoption of the above technical solution, the technical advancement achieved by this invention is as follows: This invention adds an A channel to the RGB image, and sends the dataset into... The dust concentration map obtained by the module is channel A. This dust concentration map has strong prior information, which is used to guide... The segmentation network supplements spatial information to improve segmentation performance, and... By incorporating the CBAM attention mechanism into the segmentation network, segmentation accuracy is improved, enabling not only precise dust area detection but also dust concentration classification. This invention utilizes a computer to accurately mark the contours of dust particles through monitoring video, allowing for precise estimation of dust concentration. Compared to other traditional detection methods, it has a wider range of applications and is less susceptible to environmental interference. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This involves labeling the images in the dataset using the labelme software.

[0039] Figure 3 The mask image obtained after annotation;

[0040] Figure 4 for add Improved flowchart;

[0041] Figure 5 for Improved network structure diagram with CBAM;

[0042] Figure 6 Here is the network structure diagram of CBAM;

[0043] Figure 7 These are photos of the scene, which are covered in dust.

[0044] Figure 8 This is a graph showing dust concentration. Detailed Implementation

[0045] The embodiments of the present invention will now be described in more detail with reference to the accompanying drawings and specific examples.

[0046] Combination Figure 1 The image dehazing and... The steps for detecting dust concentration levels are as follows:

[0047] Step 1: Set up a dust video monitoring system and determine the type of monitoring camera and its installation location;

[0048] (1) Select a suitable camera and a suitable location for installation. Due to the complex working environment in ports, the selection of cameras is particularly important. The selection of cameras should meet the following requirements:

[0049] (2) Seismic performance: Since the ship will vibrate during operation, it must have a certain seismic performance to avoid the image being unclear due to vibration.

[0050] (3) Waterproof performance: Since the camera is operated in an open-air environment, it must have a certain degree of waterproof performance to avoid damage caused by rain.

[0051] Because the work is conducted in an open-air environment, it must be able to operate under various ambient temperatures, generally from -30°C to +60°C.

[0052] (4) Electromagnetic interference resistance: Due to the complex working environment, there are various high-power electrical devices, which are accompanied by various electromagnetic interferences. In order to avoid electromagnetic interference affecting the camera imaging, it should have a certain electromagnetic interference resistance capability.

[0053] (5) It should be less sensitive to light, and the image should be clear whether it is under night lighting or strong daylight.

[0054] Step 2: Use the camera from Step 1 to collect images of coal dust during chute unloading under different operating conditions, label the coal dust in different scenarios, create a dataset, and obtain network training and testing samples.

[0055] Videos were captured under various conditions, including daytime and nighttime, using the camera in step 1. These videos were then converted into an image library using FreeVideo to JPG Converter. Photos from various scenes were selected from this library, along with photos of dust-free and strictly aligned images, to serve as the dataset for image dehazing. Dust samples from various conditions were then labeled using the labelme software. Figure 2 As shown, Figure 3 This is the mask image obtained after annotation.

[0056] Step 3: Classify and process dust concentration. The algorithm was specifically adjusted and optimized, including:

[0057] Step 3-1, to enhance To assess the segmentation capability and dust concentration classification, a segmentation model based on RGBA four channels is proposed, such as... Figure 4 As shown, input the image into In this process, by subtracting the resulting image from the original image and converting it to grayscale, a dust concentration image is obtained, which is the A channel image. Because... Multiple downsampling operations in the image result in the loss of much spatial information. Dust maps, rich in prior information, can not only compensate for this loss but also be stitched together with the original image to obtain an RGBA image. This RGBA image is then fed into the U-net segmentation network with an attention mechanism, significantly improving segmentation performance.

[0058] Step 3-2, add channel attention and spatial attention modules to the U-net network, because Extensive downsampling leads to the loss of much spatial information. Therefore, channel attention and spatial attention modules are added; specifically, CBAM (channel attention and spatial attention) is added after each downsampling step in the U-net network. Figure 5 As shown, Figure 5 To add CBAM to U-net, such as Figure 6 As shown, channel attention processes the input feature map F (H×W×C) through global max pooling and global average pooling, respectively, to obtain two 1×1×C feature maps. These are then fed into a two-layer neural network (MLP). The first layer has C / r neurons (r is the reduction rate) and uses ReLU activation, while the second layer has C neurons. These two layers are shared. The MLP output features are then element-wise summed and followed by a sigmoid activation operation to generate the final channel attention feature Mc. Finally, Mc is element-wise multiplied with the input feature map F to generate the input features needed by the spatial attention module. The spatial attention module then processes the feature map output by the channel attention module. This serves as the input feature map for this module. First, we perform global max pooling and global average pooling to obtain two... The feature maps are then concatenated channel by channel of the two feature maps. Then, after a... Convolution operation reduces dimensionality to one channel, i.e. The spatial attention feature, Ms, is then generated using a sigmoid function. Finally, this feature map is element-wise multiplied with the input feature map of this module to obtain the final generated feature.

[0059] Step 3-3, using The loss function with boundary weights is used, and its calculation formula is as follows:

[0060]

[0061] in It is the softmax loss function. It is the label of each pixel. The weight of the pixel

[0062]

[0063] in It is the weight that balances the proportions of the categories. This represents the distance between a pixel and its nearest dust particle. This represents the distance between a pixel and its second nearest dust particle. and It is a constant value, in the experiment. ,

[0064] Step 4: Haze and dust particles can affect image brightness. Essentially, they change the transmittance. Subtract the dust map from the output result to obtain the difference in pixel brightness values, which can be used as a characterization of dust concentration.

[0065] Step 4-1, using The defogging algorithm calculates the estimated dust concentration using the following formula:

[0066]

[0067] Where y(x) represents the pixel brightness value with spatial coordinate x in the image, J represents the restored image obtained after passing through the dehazing network, I represents the network input image which is an image with dust, A represents atmospheric light, and I and J are both grayscale images.

[0068] Step 4-2: Filter y(x) to obtain the dust concentration of the segmented region.

[0069] According to the atmospheric scattering model:

[0070]

[0071] get:

[0072]

[0073] Where t(x) is the transmittance, and since the pixel value of point J(x) is known, and t(x) is inversely proportional to the dust concentration, it can be known that y(x) is directly proportional to the dust concentration. Therefore, y(x) represents the dust concentration, and the dust concentration is graded based on the area of ​​the dust region. Figure 8 for Figure 7 The corresponding dust concentration graph.

[0074] Step 5: Set up the PyTorch deep learning environment using PyTorch 1.6.0 and TorchVision 0.7.0. Utilize the established platform and the dataset obtained in Step 2 to apply the attention mechanism added in Step 3. The segmentation network is trained to obtain the training weight file, and then OpenCV is used to perform video detection.

[0075] Step 6: Through repeated testing, find the relationship between the pixel values ​​of the dust image and the actual concentration values.

[0076] Step 6-1: Import the operation video, use the algorithm to obtain the precise location of the dust area, calculate the dust map in the area, and obtain the dust concentration in the area.

[0077] Step 6-2: Use an on-site dust concentration detector to obtain the actual dust concentration.

[0078] Step 6-3: Call up operation videos under different working conditions and test them repeatedly to find the relationship between the pixel value of the dust map and the dust concentration.

[0079] Step 7: Based on the dust concentration and the area of ​​the dusty area, make a comprehensive judgment and classify the dust concentration. Then, spray water to reduce dust according to the dust concentration and the dust concentration level. Repeatedly test to obtain the optimal relationship between the amount of water sprayed and the dust concentration.

[0080] This invention innovatively will and By combining dehazing with other deep learning detection methods, this invention adds an A channel to the RGB image, and feeds the dataset into... The dust concentration map obtained by the module is channel A. This dust concentration map has strong prior information, which is used to guide... The segmentation network supplements spatial information to improve segmentation performance, and... By incorporating the CBAM attention mechanism into the segmentation network, the segmentation accuracy is improved. This not only enables precise dust area detection but also allows for dust concentration classification. Using a trained network, dust areas and concentration levels can be automatically detected, and the amount of water sprayed for dust suppression can be controlled based on the dust concentration level. This eliminates the need for manual detection, significantly saving labor costs and water resources. It has advantages such as low detection cost, high detection accuracy, fast detection speed, and accurate and rapid dust suppression.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for dust detection and concentration grading based on prior information, characterized in that: Specifically comprising the following steps: Step 1, build a chute unloading dust video monitoring system, determine the type of monitoring camera and its installation position; Step 2, obtain the data set, including chute unloading dust images under different working conditions, and strictly paired dust-free images, and label the coal dust under different scenes; Step 3, input the data set into the Fast-DMPHN module to obtain a gray image called dust concentration map, which has prior information, and uses the prior information to guide the U-net segmentation network with attention mechanism to supplement the spatial information and improve the segmentation performance, and the original image in the data set and the dust concentration map are spliced to form an RGBA image, which is input into the U-net segmentation network with attention mechanism, and finally the dust segmentation map, i.e. the accurate position of the dust, is obtained; Specifically comprising: Step 3-1, the Fast-DMPHN module is an image defogging module, which inputs the image into the Fast-DMPHN module to obtain a gray image as a dust concentration map, which is used as prior information to guide the segmentation network for dust segmentation, and then spliced with the input image to obtain an RGBA four-channel image input into the U-net segmentation network with attention mechanism; Step 3-2, add channel attention and spatial attention models in the U-net segmentation network, i.e. add CBAM after each downsampling in the U-net segmentation network to obtain the U-net segmentation network with attention mechanism; Step 3-3, input the obtained RGBA image into the U-net segmentation network with attention, and finally obtain the dust segmentation map, i.e. the accurate position of the dust; Step 3-4, the U-net segmentation network uses a loss function with boundary weight, and the calculation formula of the loss function is: where p l(x) (x) is a softmax loss function, l : Ω → {1,..., K} is the label of each pixel point, is the weight of the pixel point; wherein is a weight balancing the class proportions, denotes the distance of the pixel point to its nearest distance dust, denotes the distance of the pixel point to its second nearest distance dust, and ω0and σ are constants; Step 4, the dust concentration map obtained by image defogging is used as the representation of dust concentration; Step 5, according to the approximate mapping relationship between the dust concentration map and the true value of the dust concentration, and the dust segmentation map, the concentration and concentration classification of the dust are realized.

2. The dust detection and concentration grading method based on prior information according to claim 1, characterized in that, The dust concentration estimation of the dust concentration map specifically comprises: Through the Fast-DMPHN module defogging algorithm, the estimated value of the dust concentration is obtained, and the calculation formula is: y(x)=I(x)-J(x) Where y(x) represents the pixel brightness value of the pixel with spatial coordinates x in the image, J represents the restored image obtained after the defogging network, I represents the network input image, i.e. the image with dust, and I and J are both gray images; According to the atmospheric scattering model: I(x)=J(x)t(x)+A(1-t(x)) We get: y(x)=(A-J(x)(1-t(x))) Where t(x) is the transmittance, A is the atmospheric light, since the pixel value of J(x) is known, and t(x) is inversely proportional to the dust concentration, it can be known that y(x) is proportional to the dust concentration, so y(x) represents the dust concentration; Filter y(x) to obtain the dust concentration of the segmentation region.

3. The dust detection and concentration grading method based on prior information according to claim 2, characterized in that, Also included is the use of Pytorch deep learning framework and the dust dataset made in step 2 to train the U-net segmentation network with prior information and attention mechanism in step 3, to obtain the final dust detection.weights weight file, and then use opencv to realize the video detection algorithm.

4. The dust detection and concentration grading method based on prior information according to claim 1, characterized in that, The dust detection and concentration grading specifically includes: The dust video monitoring system collects the video of the draw-off operation of the draw-off cylinder in real time, the U-net segmentation network with prior information and attention mechanism is used to detect each frame of image, the dust segmentation area is obtained, and the concentration value of the dust segmentation area is calculated; The real concentration value of the dust is tested on site through the laser dust concentration detector; The above two steps are repeated in different scenes to obtain the approximate mapping relationship between the real concentration value of the dust in different scenes.

5. The dust detection and concentration grading method based on prior information according to claim 4, characterized in that, When detecting the dust concentration in real time, the area of the dust segmentation area is used as important information for judging the dust concentration, the dust concentration level is comprehensively judged according to the concentration in the dust area, the dust area is watered according to the segmentation area, and the above dust concentration is used as the control standard for controlling the amount of watering.

Images