Intelligent dust recognition method and system based on machine vision

Abstract

The present application relates to the technical field of image processing, in particular to a dust intelligent identification method and system based on machine vision. The method comprises: acquiring a sequence of continuous video image frames of a target region; for any target frame in the sequence, constructing a multi-channel fusion feature input graph, which includes: RGB channel data of the target frame, wavelet transform coefficients of the target frame, and an optical flow vector field of the target frame; inputting the multi-channel fusion feature input graph into a preset dust neural network model to output a dust detection result of the target region. That is, the scheme of the present application can accurately detect the dust condition of the target region.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a dust intelligent recognition method and system based on machine vision. BACKGROUND

[0002] In the fields of industrial production (such as mines, cement plants, metallurgical workshops) and urban environment monitoring, it is crucial to monitor the concentration, particle size and other parameters of the suspended dust in the air in real time and accurately. Traditional dust monitoring technologies mainly rely on physical or chemical sampling analysis methods. For example, the filter membrane weighing method calculates the average concentration by collecting particulate matters in a certain volume of air and weighing, which has high accuracy but a cumbersome process and a long time-consuming, and cannot realize real-time online monitoring, making it difficult to deal with sudden and high-dynamic dust pollution events. Another commonly used technology is the sensor based on the light scattering principle, which can realize fast response by measuring the light intensity scattered by particulate matters after the laser beam, but its measurement range is usually limited to a single point or a small area where the sensor probe is located, and it cannot provide large-scale, multi-dimensional spatial distribution information, and is easily disturbed by environmental factors such as airflow and humidity, making it difficult to fully reflect the dust pollution situation of the entire monitoring scene.

[0003] With the rapid development of computer vision technology and artificial intelligence, using images or videos for non-contact and large-scale dust identification and monitoring has become a new research direction. Early visual methods mostly use traditional image processing techniques, which analyze the gray level, texture, edge, and other information of the image, set thresholds to segment and count dust particles, and then estimate the concentration. However, such methods are extremely sensitive to changes in light and the complexity of the background, and it is difficult to effectively distinguish dust, water mist, smoke or image noise with similar shape and color, resulting in limited recognition accuracy and robustness. In recent years, methods based on deep learning, especially convolutional neural networks (CNN), have been introduced into the field, which automatically learn the deep features of dust through data-driven methods, and to some extent, improve the accuracy of identification. However, there are still many shortcomings in the existing technology: first, most models rely only on the RGB information of a single image, ignoring the dynamic motion characteristics inherent to dust as suspended particulate matter. When there are static spots or other dynamic disturbances in the background, it is easy to cause confusion and misjudgment. Second, for models that integrate motion information (such as optical flow), the feature fusion method is mostly simple concatenation or addition, which fails to effectively exploit and utilize the deep correlation and complementarity between static texture and dynamic information, limiting the further improvement of model performance. Finally, the output form of existing methods is relatively single, usually only providing a general concentration level or a simple segmentation map, which cannot provide pixel-level refined concentration distribution, key particle size distribution, and uncertainty evaluation of the model's own prediction results at the same time, which makes the information dimension of the monitoring results insufficient, and lacks decision support basis in application scenarios that require high reliability. SUMMARY

[0004] The purpose of the present application is to propose a dust intelligent identification method and system based on machine vision, in order to solve the problem of how to more accurately detect dust in the prior art. To this end, the present application provides solutions in the following two aspects.

[0005] In the first aspect, the present application provides a dust intelligent identification method based on machine vision, comprising:

[0006] obtaining a sequence of continuous video image frames of a target region;

[0007] For any target frame in the sequence, a multi-channel fusion feature input map is constructed, which includes: the RGB channel data of the target frame, the wavelet transform coefficients of the target frame, and the optical flow vector field of the target frame;

[0008] input the multi-channel fusion feature input image into a preset dust neural network model, and output a dust detection result of the target region, the dust detection result including: a pixel-level dust concentration prediction map, in which each pixel value represents a dust concentration prediction value of a corresponding position; a dust particle size distribution vector, each dimension of which corresponds to a particulate matter proportion of a preset particle size interval; and an uncertainty mask map, in which each pixel value represents a confidence of a dust concentration prediction value of each pixel position in the target frame;

[0009] The dust neural network model includes an encoder, an attention fusion module, and a decoder; the encoder includes a texture feature extraction path and a motion feature extraction path in parallel, the texture feature extraction path is used to extract a high-frequency texture feature map from the RGB channel data and the wavelet transform coefficient, and the motion feature extraction path is used to extract a spatiotemporal motion feature map from the optical flow vector field; and the attention fusion module is used to fuse the high-frequency texture feature map and the spatiotemporal motion feature map to obtain a fusion feature.

[0010] The above scheme can fundamentally enhance the representation ability of the dust target by constructing a multi-channel fusion feature input image including RGB data, wavelet transform coefficients, and an optical flow field, so that the dust neural network model can simultaneously perceive the static appearance texture and the time sequence motion law of the dust. Meanwhile, the parallel feature extraction paths and the attention fusion module can deeply mine and utilize the internal correlation between the texture and the motion feature, and can combine the information from different sources with emphasis, thereby significantly improving the accuracy and robustness of dust recognition in the case of complex background and interference.

[0011] Optionally, the wavelet transform coefficient is obtained by: performing three-level discrete wavelet transform on the grayscale image of the target frame, and extracting three high-frequency component coefficient maps of horizontal details, vertical details, and diagonal details obtained by third-level decomposition.

[0012] The above wavelet transform coefficient can effectively capture local mutations of the image caused by dust particles.

[0013] Optionally, the optical flow vector field is obtained by: using a dense optical flow algorithm to calculate a backward optical flow field between the target frame and a previous frame, and a forward optical flow field between the target frame and a next frame; and stacking the backward optical flow field and the forward optical flow field along the channel dimension to obtain a four-channel optical flow vector field including horizontal and vertical displacement components.

[0014] The above optical flow vector field can reflect the dynamic information of the dust in the target region.

[0015] Optionally, the dust particle size distribution vector is a three-dimensional vector, and specifically includes the number proportions of particulate matters with particle sizes less than or equal to 2.5 microns, between 2.5 microns and 10 microns, and greater than 10 microns in the total particulate matters detected.

[0016] Optionally, the confidence is obtained by enabling Monte Carlo Dropout in the decoder part, performing multiple forward propagations on the fusion features of the input target frame, and calculating the variance of the dust concentration prediction values at the same pixel position in the multiple output dust concentration prediction maps.

[0017] The confidence can evaluate the accuracy of the dust neural network model.

[0018] Optionally, the texture feature extraction channel and the motion feature extraction channel both adopt an encoder structure stacked by a two-dimensional convolution layer with a convolution kernel size of 3*3, a batch normalization layer and a ReLU activation function.

[0019] Optionally, the texture feature extraction channel adopts an encoder based on a ResNet-50 architecture, and the motion feature extraction channel adopts a lightweight convolutional neural network.

[0020] Optionally, the fusion features are obtained by fusing the high-frequency texture feature maps and the spatio-temporal motion feature maps through a cross-attention mechanism.

[0021] Optionally, the method further includes the steps of denoising and enhancing each target frame in the sequence of continuous video image frames.

[0022] The above method can obtain a more accurate sequence of target frames, thereby providing effective data support for subsequent dust detection of target regions.

[0023] In a second aspect, a dust intelligent identification system based on machine vision includes:

[0024] a processor;

[0025] a memory storing computer instructions for dust intelligent identification based on machine vision, which, when executed by the processor, causes the system to perform the dust intelligent identification method based on machine vision.

[0026] The present application has the following advantages:

[0027] The present application can parallelly output three key indicators, i.e., a pixel-level concentration prediction map, a particle size distribution vector and an uncertainty mask map, by constructing a dust neural network model, thereby providing more abundant and fine monitoring information than a single concentration value or a segmentation map, introducing quantitative evaluation of the reliability of the prediction results, and providing comprehensive and highly reliable data support for precise decision-making in the fields of environmental governance and safety production. Attached Figure Description

[0028] Figure 1 The flowchart illustrating the steps of the machine vision-based intelligent dust recognition method in this embodiment is shown in the illustration.

[0029] Figure 2 The schematic diagram illustrates the structural block diagram of the machine vision-based intelligent dust recognition system in this embodiment. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0031] like Figure 1 As shown, the machine vision-based intelligent dust recognition method in this embodiment includes the following steps:

[0032] Step S1: Obtain a sequence of consecutive video image frames of the target area.

[0033] Specifically, industrial-grade high-speed cameras deployed in the target area (such as mine tunnels or cement plant workshops) continuously capture live video streams at a frame rate of 25 frames per second, and transmit the digitized video data to a back-end processing server to obtain a sequence of image frames.

[0034] Furthermore, each target frame in the sequence of continuous video image frames is subjected to denoising and enhancement processing to improve the accuracy and robustness of subsequent dust detection.

[0035] Step S2: For any target frame in the sequence, construct a multi-channel fused feature input map.

[0036] The constructed multi-channel fusion feature input map includes: RGB channel data of the target frame, wavelet transform coefficients of the target frame, and optical flow vector field of the target frame.

[0037] Specifically, for the target frame T, the data of the three color channels, R, G, and B, are extracted.

[0038] In one embodiment, a second-level discrete wavelet transform is performed on the grayscale image of the target frame T, for example, using the db4 (Daubechies 4) wavelet basis to extract its second-level high-frequency components HH (diagonal detail), HL (vertical detail), and LH (horizontal detail) as texture feature channels.

[0039] In another embodiment, the Daubechies 4 db4 wavelet basis can be used to perform a three-level discrete wavelet transform on the grayscale image of the target frame, and the three high-frequency component coefficient maps of the horizontal detail h3, vertical detail v3 and diagonal detail d3 obtained from the third-level decomposition can be extracted.

[0040] Specifically, when processing the target frame, it is first converted into a grayscale image; then, a Daubechies 4 wavelet basis with good time-frequency localization characteristics is selected to perform a three-level discrete wavelet transform on the grayscale image. After three-level decomposition, the size of the grayscale image is halved at each level, and finally the third-level low-frequency approximation component and three high-frequency detail components of equal size are obtained.

[0041] The three high-frequency components obtained above, i.e., the horizontal detail h3, the vertical detail v3 and the diagonal detail d3, can effectively capture the local mutations of the image caused by dust particles, such as edge and texture information.

[0042] In this embodiment, a dense optical flow algorithm is used to calculate the backward optical flow field between the target frame and the previous frame, and the forward optical flow field between the target frame and the next frame; the backward optical flow field and the forward optical flow field are stacked along the channel dimension to form a four-channel optical flow vector field containing horizontal and vertical displacement components.

[0043] Specifically, the dense optical flow algorithm can be PWC-Net algorithm or Farnebäck algorithm. Taking the Farnebäck algorithm as an example, the displacement vector of each pixel point in the target frame is calculated to obtain the backward optical flow field from the Nth frame to the N-1th frame, and the horizontal and vertical displacements of each pixel point are obtained to form a two-channel backward optical flow map. Similarly, the forward optical flow field from the Nth frame to the N+1th frame is calculated to obtain a two-channel forward optical flow map. The two two-channel optical flow maps are spliced along the channel dimension to form a four-channel optical flow vector field.

[0044] The four-channel optical flow vector field obtained above can comprehensively capture the dynamic information of the dust in the target frame, providing rich data containing past and future motion trends for subsequent motion feature analysis.

[0045] Step S3: inputting the multi-channel fusion feature map into a preset dust neural network model to output a dust detection result, so as to realize dust detection of the target region.

[0046] The dust neural network model comprises an encoder, an attention fusion module and a decoder.

[0047] The encoder comprises a texture feature extraction path and a motion feature extraction path connected in parallel.

[0048] The input of the texture feature extraction path is six-channel data composed of RGB and wavelet transform coefficients, and deep texture information is extracted through multi-level convolution and residual connection. The motion feature extraction path is used to process the optical flow vector field to capture the motion pattern and trajectory of the dust.

[0049] Specifically, for the texture feature extraction path, its input is a six-channel data, specifically: RGB three-channel image data and three high-frequency wavelet coefficient maps of h3, v3, d3. The input of the motion feature extraction path is a four-channel optical flow vector field.

[0050] In an embodiment, the texture feature extraction path adopts an encoder based on the ResNet-50 architecture; the motion feature extraction path adopts a lightweight convolutional neural network, for example, a structure containing three consecutive convolutional blocks. The two paths independently extract features from their respective inputs to generate high-frequency texture feature maps and spatiotemporal motion feature maps of the same size.

[0051] In another embodiment, the texture feature extraction path and the motion feature extraction path can also each adopt an encoder structure stacked by two-dimensional convolutional layers with a convolution kernel size of 3x3, batch normalization layers, and ReLU activation functions.

[0052] The two parallel feature extraction paths described above share similar encoder structures to ensure the consistency of feature scales. Through the dual-path design, the dust neural network model can simultaneously and independently learn the static appearance texture features and dynamic spatiotemporal motion features of dust, laying a foundation for subsequent precise fusion analysis.

[0053] The attention fusion module is configured to dynamically calculate attention weights according to the mutual dependency relationship between the high-frequency texture feature map and the spatiotemporal motion feature map, and perform weighted fusion on the two types of feature maps to generate a fusion feature.

[0054] Specifically, the process of obtaining the fusion feature is as follows:

[0055] The high-frequency texture feature map and the spatiotemporal motion feature map are spliced along the channel dimension, and the spliced feature map is input into at least one self-attention module. The self-attention module recalibrates the feature response by calculating the dependency relationship between different spatial positions and channels in the feature map, generates a weight for each pixel pair and channel pair in the feature map, and the weight represents the correlation strength between them. For example, significant motion features in a certain region of the image will make the corresponding static texture features obtain higher attention.

[0056] For example, the size of the high-frequency texture feature map and the spatiotemporal motion feature map is 64x64x256, the attention fusion module splices the two feature maps along the channel dimension to form a feature map with a size of 64x64x512, and the feature map is input into a self-attention module to obtain the weight of each pixel point in the feature map.

[0057] The self-attention module described above can adopt a mutual cross-attention mechanism (MCA), a channel attention mechanism, etc.

[0058] Take cross attention mechanism as an example, the process of obtaining the fusion feature of high-frequency texture feature map and spatio-temporal motion feature map is introduced as follows:

[0059] The high-frequency texture feature map is taken as the query, the spatio-temporal motion feature map is taken as the key and the value, and an attention map is calculated through dot product, scaling and Softmax function, which represents the attention degree of each texture feature position to all motion features. The spatio-temporal motion feature map is weighted and summed by using the attention map to generate motion-aware texture features. Conversely, the spatio-temporal motion feature map is taken as the query, the high-frequency texture feature map is taken as the key and the value, and similar operations are performed to generate texture-aware motion features. The two features enhanced by mutual weighting are added element by element or concatenated in the channel to obtain the final fusion feature.

[0060] The above attention fusion model can automatically and selectively enhance the feature response of the region where the texture and motion information are highly correlated, and suppress irrelevant or noise information, thereby generating a fusion feature with higher information density and stronger discrimination.

[0061] The decoder in this embodiment adopts a similar structure to the U-Net decoder, and gradually restores the fusion feature to the original input resolution through a series of upsampling or transposed convolution operations.

[0062] Among them, the decoder has three parallel output heads: the first head, i.e. the concentration prediction head, directly outputs a single-channel dust concentration prediction map, and the floating-point value of each pixel point corresponds to the concentration value of PM2.5 or TSP, for example, in micrograms per cubic meter; the second head, i.e. the particle size distribution head, performs global average pooling on the fusion feature to obtain a feature vector, and then sends it to a fully connected layer with M output nodes, where M is the number of preset particle size intervals, such as 0-2.5 microns, 2.5-10 microns, 10-100 microns, etc., and outputs the proportion of each interval by the Softmax function; the third head, i.e. the uncertainty prediction head, is similar in structure to the concentration prediction head and outputs a single-channel uncertainty mask map in parallel, whose pixel value quantifies the uncertainty of the model prediction.

[0063] Among them, the dust particle size distribution vector is a three-dimensional vector, and each dimension respectively represents the number proportion of particulate matter with particle size less than or equal to 2.5 microns, between 2.5 microns and 10 microns, and greater than 10 microns in the total particulate matter detected.

[0064] Exemplarily, the decoder output is a normalized three-dimensional vector, respectively, seventy percent, twenty percent, and ten percent. The three components correspond to three preset particle size intervals, respectively. The first component is 0.7, which indicates that seventy percent of all dust particles detected in the current frame have an aerodynamic equivalent diameter less than or equal to 2.5 microns, which can be regarded as fine particulate matter. The second component is 0.2, which indicates that twenty percent of the particles have a particle size between 2.5 microns and 10 microns. The third component is 0.1, which represents that the proportion of coarse particles with a particle size greater than 10 microns is ten percent. The sum of the three components is 1.

[0065] The dust particle size distribution vector described above reveals the situation of dust particles of different sizes in the environment, providing fine data support for air quality assessment.

[0066] wherein the uncertainty mask map is obtained by repeatedly performing forward propagation on the same input multiple times with Monte Carlo Dropout technique enabled in the decoder part, and calculating the variance of the pixel-level dust concentration prediction maps at each pixel position of the multiple outputs.

[0067] wherein the Monte Carlo Dropout technique is a deep learning inference method combining Dropout regularization technique with Monte Carlo method, used to quantify the prediction uncertainty of neural network models. Traditional Dropout prevents overfitting by randomly discarding neurons during training, but usually turns off Dropout during testing to generate deterministic predictions. In contrast, MC Dropout keeps Dropout active during the testing phase, simulates different possible implementations of the model by performing multiple forward propagations (each time randomly discarding different neurons), and thus estimates the distribution of predictions, rather than a single point estimate.

[0068] Specifically, in order to quantify the reliability of the model prediction results, Monte Carlo Dropout technique is adopted in the inference phase. In conventional inference, the Dropout layer is turned off to ensure the determinacy of the prediction. Here, the Dropout layer is kept active in the decoder part, and 50 independent forward propagations are performed on the same input sample, for example, the same frame of image. Since the neurons that are randomly inactivated are different in each propagation, 50 slightly different pixel-level dust concentration prediction maps will be obtained. For each pixel position in the image, the variance of the 50 prediction values at the position is calculated, and the generated variance map is the uncertainty mask map. The areas with higher brightness in the map, i.e. the pixels with larger variance, represent lower confidence of the model's prediction results at these positions, and vice versa.

[0069] It should be noted that before dust detection of the target region of the scene to be identified, the dust neural network model needs to be trained to obtain the trained dust neural network model.

[0070] The specific training process is as follows:

[0071] First, a historical training set is obtained. The training set includes an input image set and a label.

[0072] The image set is a multi-channel fusion feature input image of each frame image in the historical image frame sequence in the detection region; the label includes the dust concentration of each pixel position in each frame image, the number ratio of the size of particulate matter in the detection region, and 0.

[0073] Second, the dust neural network model is trained using the historical training set to obtain the trained dust neural network model.

[0074] Since the training process is prior art, it will not be described in detail here.

[0075] Further, the multi-channel fusion feature input image of the target region of the scene to be identified is input into the trained dust neural network model to output a dust detection result.

[0076] According to the output dust detection result in the embodiment, it can be determined whether the dust in the target region is qualified, and the staff can be reminded in time when it is not qualified.

[0077] The scheme of the present application can combine information from different sources with emphasis, thereby significantly improving the accuracy and robustness of dust identification in the presence of complex background and interference, and achieving the purpose of dust detection in the target region.

[0078] The present application also provides a dust intelligent identification system based on machine vision. As shown in Figure 2 The system includes a processor and a memory, and the memory stores computer program instructions, which, when executed by the processor, implement the dust intelligent identification method based on machine vision according to the present application.

[0079] The system also includes a communication bus and a communication interface and other components familiar to those skilled in the art, the settings and functions of which are known in the art, and therefore will not be described here.

[0080] In this description, the term "application" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "memory" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "computer-readable medium" means any tangible medium that stores, communicates, or otherwise provides data that can be used by an instruction execution system, apparatus or device. The computer-readable medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as, for example, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described in this description can be implemented by computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0081] In the description of the present description, the meaning of "a plurality of" is at least two, for example, two, three or more, and the like, unless otherwise explicitly specified.

[0082] Although the present description has shown and described a number of embodiments of the present application, it will be apparent to those skilled in the art that many modifications, variations, and alternatives to the embodiments described herein can be made in light of the teachings herein.

Claims

1. A machine vision-based intelligent dust recognition method, characterized in that, include: Obtain a sequence of consecutive video image frames for the target region; For any target frame in the sequence, a multi-channel fusion feature input map is constructed, which includes: RGB channel data of the target frame, wavelet transform coefficients of the target frame, and optical flow vector field of the target frame; The multi-channel fused feature input map is input into a preset dust neural network model, and the dust detection results of the target area are output. The dust detection results include: a pixel-level dust concentration prediction map, in which each pixel value represents the predicted dust concentration value at the corresponding position; a dust particle size distribution vector, in which each dimension corresponds to the proportion of particles in a preset particle size range; and an uncertainty mask map, in which each pixel value represents the confidence level of the predicted dust concentration value at each pixel position in the target frame. The dust neural network model includes an encoder, an attention fusion module, and a decoder. The encoder includes a parallel texture feature extraction path and a motion feature extraction path. The texture feature extraction path is used to extract high-frequency texture feature maps from the RGB channel data and wavelet transform coefficients, and the motion feature extraction path is used to extract spatiotemporal motion feature maps from the optical flow vector field. The attention fusion module is used to fuse the high-frequency texture feature maps and the spatiotemporal motion feature maps to obtain fused features.

2. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, The process of obtaining the wavelet transform coefficients is as follows: perform a three-level discrete wavelet transform on the grayscale image of the target frame, and extract the three high-frequency component coefficient maps of horizontal details, vertical details and diagonal details obtained from the third-level decomposition.

3. The machine vision-based intelligent dust identification method according to claim 2, characterized in that, The process of obtaining the optical flow vector field is as follows: using a dense optical flow algorithm, the backward optical flow field between the target frame and the previous frame, and the forward optical flow field between the target frame and the next frame are calculated respectively; the backward optical flow field and the forward optical flow field are stacked along the channel dimension to obtain a four-channel optical flow vector field containing horizontal and vertical displacement components.

4. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, The dust particle size distribution vector is a three-dimensional vector, specifically including the proportion of particles with a diameter less than or equal to 2.5 micrometers, between 2.5 micrometers and 10 micrometers, and larger than 10 micrometers in the total detected particles.

5. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, The confidence level is calculated by enabling Monte Carlo Dropout in the decoder section, performing multiple forward propagations on the fused features of the input target frame, and calculating the variance of the dust concentration prediction values ​​at the same pixel position in the multiple output dust concentration prediction maps.

6. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, Both the texture feature extraction pathway and the motion feature extraction pathway adopt an encoder structure consisting of a stacked two-dimensional convolutional layer with a kernel size of 3×3, a batch normalization layer, and a ReLU activation function.

7. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, The texture feature extraction path uses an encoder based on the ResNet-50 architecture; the motion feature extraction path uses a lightweight convolutional neural network.

8. The intelligent dust recognition method based on machine vision according to claim 1, characterized in that, The fusion feature is obtained by fusing high-frequency texture feature maps and spatiotemporal motion feature maps using a cross-attention mechanism.

9. The machine vision-based intelligent dust identification method according to claim 1, characterized in that, Also includes: The steps of denoising and enhancing each target frame in the sequence of continuous video image frames.

10. A machine vision-based intelligent dust recognition system, characterized in that, include: processor; A memory storing computer instructions for machine vision-based intelligent dust identification, which, when executed by the processor, cause the system to perform the machine vision-based intelligent dust identification method according to any one of claims 1-9.

Images