Atomization state recognition method and device of atomization nozzle and training method of atomization state recognition model

Abstract

The disclosure provides an atomization state recognition method and device of an atomizing nozzle and a training method of an atomization state recognition model, which can be applied to the technical fields of image processing and computer vision. The atomization state recognition method of the atomizing nozzle comprises: performing droplet edge enhancement processing on an original atomization image generated by the atomizing nozzle in a production process to obtain a target atomization image without imaging noise; performing multi-dimensional asymmetric convolution processing on the target atomization image to obtain multi-dimensional atomization features; performing weight mask processing on the multi-dimensional atomization features based on a spatial attention mechanism to obtain spray profile features with droplet boundary information and spray cone structure information, and performing weight fusion processing on the spray profile features based on a double-head attention mechanism to obtain atomization fusion features; and performing classification processing on the atomization fusion features to obtain an atomization state recognition result of the atomizing nozzle.

Description

Technical Field

[0001] This disclosure relates to the fields of image processing and computer vision technology, and specifically to a method, apparatus, and training method for a fogging state recognition model of a fogging nozzle. Background Technology

[0002] Spraying technology is widely used in industrial production, agricultural irrigation, environmental management, and medical fields. The quality of its spraying directly affects the atomization effect and the stability and uniformity of subsequent processes. In practical applications, nozzles are prone to abnormalities such as dripping, leakage, clogging, or uneven spraying due to long-term operation or external environmental influences. This leads to a decrease in atomization effect and, in severe cases, may cause system performance instability or product quality problems.

[0003] Traditional nozzle spray status detection mainly relies on manual observation or simple sensor detection. These methods are limited by personnel experience, ambient lighting, and changes in spray characteristics, often making it difficult to achieve accurate and stable monitoring. Furthermore, manual detection is inefficient and lacks real-time performance, failing to meet the demands of modern industrial processes for automated and intelligent monitoring.

[0004] With the development of computer vision and image processing technologies, image-based spray detection methods have gradually attracted attention. These methods can reflect the spray state of the nozzle relatively intuitively by acquiring and analyzing spray images. However, existing technologies mostly use simple threshold judgment methods based on image grayscale or morphological features, which have poor robustness and are sensitive to factors such as ambient lighting and changes in atomized particle size, making it difficult to accurately identify complex spray states. Summary of the Invention

[0005] In view of the above problems, this disclosure provides a method, device and training method for atomization state recognition model of atomizing nozzle.

[0006] According to the first aspect of this disclosure, a method for identifying the atomization state of an atomizing nozzle is provided, comprising: performing droplet edge enhancement processing on the original atomization image generated by the atomizing nozzle during the production process to obtain a target atomization image with imaging noise removed, wherein the imaging noise includes imaging noise generated by the light fog effect and imaging noise generated by water vapor particles during the production process; performing multi-dimensional asymmetric convolution processing on the target atomization image to obtain multi-dimensional atomization features, wherein the multi-dimensional atomization features are used to characterize the features of the atomization region in the target atomization image from the horizontal and vertical dimensions; performing weight masking processing on the multi-dimensional atomization features based on a spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information, and performing weight fusion processing on the spray contour features based on a dual-head attention mechanism to obtain atomization fusion features; and performing classification processing on the atomization fusion features to obtain the atomization state identification result of the atomizing nozzle.

[0007] According to embodiments of this disclosure, the above-mentioned droplet edge enhancement processing of the original atomized image generated by the atomizing nozzle during the production process to obtain a target atomized image with removed imaging noise includes: obtaining a gray-level histogram of the original atomized image by counting the number of pixels appearing at each gray level in the original atomized image; calculating a gray-level cumulative distribution function using the gray-level histogram; and remapping the gray-level values ​​in the original atomized image using the gray-level cumulative distribution function to achieve droplet edge enhancement in the original atomized image, thereby obtaining a target atomized image with removed imaging noise.

[0008] According to embodiments of this disclosure, the above-described process of enhancing the droplet edges of the original atomized image generated during the production process of the atomizing nozzle to obtain a target atomized image with removed imaging noise includes: calculating the spatial distance between pixels in the original atomized image to obtain spatial weights, and calculating the similarity between pixel values ​​in the original atomized image to obtain pixel value domain weights; performing calculations on the spatial weights and pixel value domain weights to obtain fusion weights; and using the fusion weights to perform pixel value weighted averaging on the original atomized image to enhance the droplet edges in the original atomized image, thereby obtaining a target atomized image with removed imaging noise.

[0009] According to embodiments of this disclosure, the above-mentioned multi-dimensional asymmetric convolution processing of the target fogged image to obtain multi-dimensional fogging features includes: performing asymmetric convolution processing on the target fogged image in the horizontal dimension to obtain horizontal fogging features, and performing asymmetric convolution processing on the target fogged image in the vertical dimension to obtain vertical fogging features; and using structural reparameterization technology, weightedly fusing the horizontal fogging features and the vertical fogging features to obtain multi-dimensional fogging features.

[0010] According to embodiments of this disclosure, the above-mentioned weight masking processing of multi-dimensional atomization features based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information includes: performing global average pooling processing on the multi-dimensional atomization features to obtain globally average pooled multi-dimensional atomization features, and performing global max pooling processing on the multi-dimensional atomization features to obtain globally max pooled multi-dimensional atomization features; performing convolution activation processing on the channel dimension of the globally average pooled multi-dimensional atomization features and the globally max pooled multi-dimensional atomization features to obtain a spatial attention weight map; and enhancing the droplet boundary and spray cone structure of the atomization region by performing weight masking processing on the spatial attention weight map and the multi-dimensional atomization features to obtain spray contour features.

[0011] According to embodiments of this disclosure, the above-mentioned weighted fusion processing of the spray contour features based on a dual-head attention mechanism to obtain atomization fusion features includes: enhancing the spray contour features in the channel dimension based on a global attention head mechanism to obtain global spray features, wherein the spray shape includes the symmetry of the spray and the diffusion range of the spray; enhancing the spray contour features in the local spray characteristics based on a local attention head mechanism to obtain local spray features, wherein the local spray characteristics include droplet shape, spray angle and spray splash state; and fusing the global spray state and the local spray state based on a weighted weighted processing of a spatial attention weight map to obtain atomization fusion features.

[0012] According to a second aspect of this disclosure, an atomization state recognition device for an atomizing nozzle is provided, comprising: an imaging noise removal module, used to perform droplet edge enhancement processing on the original atomization image generated by the atomizing nozzle during the production process to obtain a target atomization image with imaging noise removed, wherein the imaging noise includes imaging noise generated by the light fog effect and imaging noise generated by water vapor particles during the production process of the atomizing nozzle; an asymmetric convolution module, used to perform multi-dimensional asymmetric convolution processing on the target atomization image to obtain multi-dimensional atomization features, wherein the multi-dimensional atomization features are used to characterize the features of the atomization region in the target atomization image from the horizontal and vertical dimensions; an attention mechanism module, used to perform weight mask processing on the multi-dimensional atomization features based on a spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information, and to perform weight fusion processing on the spray contour features based on a dual-head attention mechanism to obtain atomization fusion features; and an atomization state classification module, used to perform classification processing on the atomization fusion features to obtain the atomization state recognition result of the atomizing nozzle.

[0013] According to a third aspect of this disclosure, a training method for a fogging state recognition model is provided. The fogging state recognition model includes an asymmetric convolution module, a spatial attention module, a dual-head attention module, and a classification module. The method includes: performing multi-dimensional asymmetric convolution processing on fogging image samples using the asymmetric convolution module to obtain multi-dimensional fogging feature samples. The fogging image samples are obtained by preprocessing a fogging video of a fogging nozzle. The multi-dimensional fogging features are used to characterize the fogging region in the target fogging image from both horizontal and vertical dimensions. The method further includes performing weight masking processing on the multi-dimensional fogging feature samples based on a spatial attention mechanism using the spatial attention module to obtain a fogging feature sample with fog droplet edges. The process involves: 1) collecting spray contour feature samples containing boundary information and spray cone structure information; 2) performing weighted fusion processing on the spray contour feature samples using a dual-head attention module to obtain atomization fusion feature samples; 3) classifying the atomization fusion feature samples using a classification module to obtain the atomization state prediction result of the atomizing nozzle; 4) processing the atomization state prediction result and the label values ​​of the atomization image samples using a preset cross-entropy loss function to obtain the cross-entropy loss value, and then using the cross-entropy loss value to optimize the parameters of the atomization state recognition model; and 5) iteratively performing data processing and parameter optimization operations on the atomization state recognition model until the preset training conditions are met, resulting in a trained atomization state recognition model.

[0014] According to embodiments of this disclosure, the aforementioned atomized image samples are obtained by preprocessing the atomization video of the atomizing nozzle, including: decoding the atomization video to obtain a sequence of atomized image frames; filtering the sequence of atomized image frames according to preset filtering criteria to obtain a target atomized frame sequence, wherein the preset filtering criteria include image brightness, image contrast, and gamma distribution information of the image; and using a preset algorithm to enhance the droplet edges of the target atomized frame sequence to obtain atomized image samples with removed imaging noise, wherein the preset algorithm includes a histogram equalization algorithm and a bilateral filtering algorithm.

[0015] According to embodiments of this disclosure, the above-mentioned multi-dimensional asymmetric convolution processing of fogged image samples using an asymmetric convolution module to obtain multi-dimensional fogged feature samples includes: performing asymmetric convolution processing on the fogged image samples in the horizontal dimension using the horizontal convolution units of the asymmetric convolution module to obtain horizontal dimension fogged feature samples; performing asymmetric convolution processing on the fogged image samples in the vertical dimension using the vertical convolution units of the asymmetric convolution module to obtain vertical dimension fogged feature samples; and using a structure reparameterization technique, weightedly fusing the horizontal fogged feature samples and the vertical dimension fogged feature samples using the asymmetric convolution module to obtain multi-dimensional fogged feature samples.

[0016] A fourth aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0017] The fifth aspect of this disclosure also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0018] A sixth aspect of this disclosure also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] The atomization state recognition method for atomizing nozzles disclosed herein effectively removes imaging noise through edge enhancement processing, providing high-quality image data for subsequent atomization state recognition. It employs asymmetric convolution processing to capture detailed information of the atomization region from multiple dimensions. Combining spatial attention and dual-head attention mechanisms, it highlights key features and improves recognition accuracy. It can effectively extract droplet boundaries and spray cone structure information, enhancing the model's understanding of complex atomization states. Furthermore, the method disclosed herein can efficiently extract and analyze key features from spray images under complex backgrounds and variable operating conditions, achieving accurate identification and monitoring of spray state, atomization effect, and abnormal conditions. The method disclosed herein can be widely applied in industrial automated production, intelligent detection, and spray process optimization, providing real-time monitoring and decision support for related processes. Attached Figure Description

[0020] The foregoing contents, as well as other objects, features, and advantages of this disclosure, will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 An application scenario diagram of the atomization state recognition method for an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0022] Figure 2 A flowchart of a method for identifying the atomization state of an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0023] Figure 3 An architecture diagram of model training and inference according to an embodiment of the present disclosure is shown.

[0024] Figure 4 A comparative schematic diagram of horizontal convolution and standard convolution according to an embodiment of the present disclosure is shown.

[0025] Figure 5The training and inference processes of an asymmetric convolution module according to an embodiment of this disclosure are illustrated.

[0026] Figure 6 The RepVGG_ACB network architecture according to an embodiment of this disclosure is shown.

[0027] Figure 7 The training and inference phases of the RepVGG_ACB network according to an embodiment of this disclosure are shown.

[0028] Figure 8 A structural diagram of a spatial attention module and a dual-head attention module according to embodiments of the present disclosure is shown.

[0029] Figure 9 A typical atomization state diagram according to an embodiment of the present disclosure is shown.

[0030] Figure 10 A structural block diagram of an atomization state recognition device for an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0031] Figure 11 A block diagram of an electronic device suitable for implementing a fogging state recognition method according to an embodiment of the present disclosure is shown. Detailed Implementation

[0032] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0034] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0035] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0036] Existing atomization state recognition technologies for atomizing nozzles suffer from several drawbacks in complex atomization scenarios, including sensitivity to changes in atomization image direction, difficulty in focusing on key spray areas, and insufficient ability to simultaneously perceive both global morphology and local anomalies. To address at least one of these issues, this disclosure provides a method, apparatus, and training method for an atomization state recognition model for atomizing nozzles. This method can more accurately distinguish between normal spraying and abnormal states such as dripping and leakage, providing technical support for the intelligent monitoring of atomization state recognition, particularly for spraying conditions within a spray drum.

[0037] Figure 1 An application scenario diagram of the atomization state recognition method for an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0038] like Figure 1 As shown, application scenario 100 according to this embodiment may include image processing and computer vision, etc. Network 104 is used as a medium to provide a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. Network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0039] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0040] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0041] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0042] It should be noted that the atomization state identification method for the atomizing nozzle provided in this embodiment can generally be executed by server 105. Correspondingly, the atomization state identification device for the atomizing nozzle provided in this embodiment can generally be located in server 105. The atomization state identification method for the atomizing nozzle provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the atomization state identification device for the atomizing nozzle provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0043] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0044] The following will be based on Figure 1 The described scene, through Figures 2-9 The atomization state recognition method of the atomizing nozzle according to the disclosed embodiments will be described in detail.

[0045] Figure 2 A flowchart of a method for identifying the atomization state of an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0046] like Figure 2 As shown, the atomization state identification method of the atomizing nozzle in this embodiment includes operations S210 to S240.

[0047] In operation S210, the original atomized image generated by the atomizing nozzle during the production process is subjected to droplet edge enhancement processing to obtain a target atomized image with imaging noise removed. The imaging noise includes imaging noise generated by the light fog effect of the atomizing nozzle during the production process and imaging noise generated by water vapor particles.

[0048] The aforementioned atomizing nozzle, such as an industrial atomizing nozzle, produces atomized images that, since these images are collected from an industrial site, may undergo unexpected directional changes (such as vertical or horizontal flipping) due to factors such as the camera's installation angle or reinstallation after maintenance.

[0049] In actual production processes, atomizing nozzles are typically located in high-temperature and high-humidity environments. During the spraying of liquid substances, Mie scattering and other factors affect the imaging of the atomized image. Simultaneously, water vapor particles near the nozzle also scatter light, further impacting the atomized image. Therefore, denoising of the original atomized image is necessary.

[0050] In operation S220, a multi-dimensional asymmetric convolution process is performed on the target fogged image to obtain multi-dimensional fogging features. These multi-dimensional fogging features are used to characterize the fogged region in the target fogged image from both horizontal and vertical dimensions.

[0051] The target fogged image is processed using asymmetric convolution, for example, by using the asymmetric convolution module of a trained fogged state recognition model. Because existing convolutional neural networks are sensitive to changes in the direction of fogged images in complex fogged scenes, struggle to focus on key spray areas, and lack the ability to simultaneously perceive global morphology and local anomalies, this disclosure provides a fogged state recognition model. This model employs a feature extraction framework that deeply integrates asymmetric convolution, structural reparameterization, and hierarchical attention mechanisms. The fogged state recognition model provided in this disclosure specifically integrates and adaptively improves existing technologies to address the inherent characteristics of fogged image data (e.g., geometric symmetry, spatial heterogeneity, and multi-scale discriminability), forming a highly specialized and inference-efficient feature extraction backbone network specifically designed to improve the robustness and accuracy of fogged state recognition.

[0052] The aforementioned multi-dimensional atomization features mainly include horizontal and vertical dimensions. Since the identification of nozzle atomization status is mainly focused on the atomization area near the nozzle, it is necessary to segment the atomization area from multiple angles of the atomization image.

[0053] In operation S230, the multi-dimensional atomization features are processed by weight masking based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information. The spray contour features are then processed by weight fusion based on dual-head attention mechanism to obtain atomization fusion features.

[0054] Attention mechanisms are applied to multi-dimensional fogging features, such as the attention mechanism module of a trained fogging state recognition model, mainly including a spatial attention mechanism module and a dual-head attention mechanism module. The spatial attention mechanism module effectively enhances the features of the fogged region in the fogged image, allowing fogging state recognition to focus on the spray area rather than other background areas. The dual-head attention mechanism enables fogging state recognition to consider both the macroscopic morphology and microscopic anomalies of the fogged region, thereby improving the accuracy of fogging state recognition.

[0055] In operation S240, the atomization fusion features are classified and processed to obtain the atomization state recognition result of the atomization nozzle.

[0056] The atomization status includes six categories: spray angle too small, spray angle too large, leakage, blurred atomization image, normal spray, and no spray image.

[0057] If the classification result is normal atomization, the operating system associated with the atomizing nozzle continues to operate and continuously monitor the nozzle status; if the classification result is abnormal, the alarm module is immediately triggered, and the abnormality type and time information are displayed on the control interface of the operating system.

[0058] The operating system uploads the abnormal atomization identification results to the upper-level MES (Manufacturing Execution System) or equipment management system to realize production tracking and maintenance prompts for the nozzle atomization status.

[0059] Based on the alarm information, the relevant operators modified the nozzle parameters to ensure the uniformity of atomization and the stability of spray during the feeding process.

[0060] Through the above implementation methods, high-precision, low-latency real-time identification can be achieved during the atomization process of the feeding nozzle, with an identification accuracy rate of over 96%. This provides support for the timely detection and handling of nozzle atomization abnormalities, and improves the stability of the feeding process and the consistency of product quality.

[0061] The atomization state recognition method for atomizing nozzles disclosed herein effectively removes imaging noise through edge enhancement processing, providing high-quality image data for subsequent atomization state recognition. It employs asymmetric convolution processing to capture detailed information of the atomization region from multiple dimensions. Combining spatial attention and dual-head attention mechanisms, it highlights key features and improves recognition accuracy. It can effectively extract droplet boundaries and spray cone structure information, enhancing the model's understanding of complex atomization states. Furthermore, the method disclosed herein can efficiently extract and analyze key features from spray images under complex backgrounds and variable operating conditions, achieving accurate identification and monitoring of spray state, atomization effect, and abnormal conditions. The method disclosed herein can be widely applied in industrial automated production, intelligent detection, and spray process optimization, providing real-time monitoring and decision support for related processes.

[0062] According to embodiments of this disclosure, the above-mentioned droplet edge enhancement processing of the original atomized image generated by the atomizing nozzle during the production process to obtain a target atomized image with removed imaging noise includes: obtaining a gray-level histogram of the original atomized image by counting the number of pixels appearing at each gray level in the original atomized image; calculating a gray-level cumulative distribution function using the gray-level histogram; and remapping the gray-level values ​​in the original atomized image using the gray-level cumulative distribution function to achieve droplet edge enhancement in the original atomized image, thereby obtaining a target atomized image with removed imaging noise.

[0063] The embodiments disclosed above, for example, employ a histogram equalization algorithm to denoise a target hazy image. The histogram equalization algorithm first scans the original hazy image, counting the number of pixels corresponding to each gray level (0-255) to form a gray-level histogram. Based on the gray-level histogram, a cumulative distribution function (CDF) is calculated, which represents the proportion of pixels with a gray value less than or equal to a certain value among the total number of pixels. The calculated CDF is then used to remap the gray values ​​of each pixel in the original image, mapping the original gray values ​​to a new gray level. This gray-level remapping enhances image contrast, making the edges of the fog droplets clearer, while simultaneously removing noise interference during the imaging process.

[0064] The embodiments provided in this disclosure process images based on their own grayscale distribution, without the need for pre-setting parameters, and are applicable to fogged images under different lighting conditions. Through the principle of histogram equalization, the contrast between fog droplets and the background is effectively improved, making edge features more prominent. While enhancing the edges, random noise in the image can be effectively suppressed, improving image quality, and without losing useful information from the original image.

[0065] According to embodiments of this disclosure, the above-described process of enhancing the droplet edges of the original atomized image generated during the production process of the atomizing nozzle to obtain a target atomized image with removed imaging noise includes: calculating the spatial distance between pixels in the original atomized image to obtain spatial weights, and calculating the similarity between pixel values ​​in the original atomized image to obtain pixel value domain weights; performing calculations on the spatial weights and pixel value domain weights to obtain fusion weights; and using the fusion weights to perform pixel value weighted averaging on the original atomized image to enhance the droplet edges in the original atomized image, thereby obtaining a target atomized image with removed imaging noise.

[0066] The embodiments of this disclosure, for example, employ a bilateral filtering algorithm to denoise a target hazy image. These embodiments effectively preserve image edge features and avoid edge blurring during denoising; they also exhibit good suppression of common noises such as Gaussian noise and salt-and-pepper noise; and they achieve adaptive filtering by calculating weights based on spatial distance and pixel similarity, with high computational efficiency, making them suitable for real-time industrial processing scenarios.

[0067] According to embodiments of this disclosure, the above-mentioned multi-dimensional asymmetric convolution processing of the target fogged image to obtain multi-dimensional fogging features includes: performing asymmetric convolution processing on the target fogged image in the horizontal dimension to obtain horizontal fogging features, and performing asymmetric convolution processing on the target fogged image in the vertical dimension to obtain vertical fogging features; and using structural reparameterization technology, weightedly fusing the horizontal fogging features and the vertical fogging features to obtain multi-dimensional fogging features.

[0068] The aforementioned extraction of multi-dimensional features, for example, utilizes the asymmetric convolution module of a trained atomization state recognition model. The horizontal convolution operation of the asymmetric convolution module extracts features of the horizontal structure of the target atomized image (e.g., the symmetry of the spray); while the vertical convolution operation extracts features of the vertical spatial structure of the target atomized image (e.g., the vertical diffusion range of the spray). Through feature extraction in different directions, key features such as droplet boundaries and spray contours in the atomized image can be extracted.

[0069] According to embodiments of this disclosure, the above-mentioned weight masking processing of multi-dimensional atomization features based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information includes: performing global average pooling processing on the multi-dimensional atomization features to obtain globally average pooled multi-dimensional atomization features, and performing global max pooling processing on the multi-dimensional atomization features to obtain globally max pooled multi-dimensional atomization features; performing convolution activation processing on the channel dimension of the globally average pooled multi-dimensional atomization features and the globally max pooled multi-dimensional atomization features to obtain a spatial attention weight map; and enhancing the droplet boundary and spray cone structure of the atomization region by performing weight masking processing on the spatial attention weight map and the multi-dimensional atomization features to obtain spray contour features.

[0070] The above embodiments of this disclosure, for example, employ a spatial attention module of a trained fogging state recognition model: by masking multi-dimensional fogging features using a spatial attention weight map, the spatial regions most relevant to the fogging state can be highlighted, such as the dense spray core area near the nozzle exit and the cone-shaped area of ​​stable spray, while suppressing the response of irrelevant backgrounds (such as equipment supports and background panels). This allows subsequent computational resources of the network to be concentrated on the image regions that are truly useful for discrimination, directly addressing the challenge of the non-fixed spatial distribution of target regions in fogged images.

[0071] According to embodiments of this disclosure, the above-mentioned weighted fusion processing of the spray contour features based on a dual-head attention mechanism to obtain atomization fusion features includes: enhancing the spray contour features in the channel dimension based on a global attention head mechanism to obtain global spray features, wherein the spray shape includes the symmetry of the spray and the diffusion range of the spray; enhancing the spray contour features in the local spray characteristics based on a local attention head mechanism to obtain local spray features, wherein the local spray characteristics include droplet shape, spray angle and spray splash state; and fusing the global spray state and the local spray state based on a weighted weighted processing of a spatial attention weight map to obtain atomization fusion features.

[0072] The embodiments disclosed above, for example, employ a dual-head attention module of a trained atomization state recognition model. This dual-head attention module collaboratively models macroscopic morphology and microscopic anomalies. The global attention head focuses on modeling the channel relationships of the entire feature map, capturing and enhancing feature channels related to the overall spray morphology, such as overall patterns reflecting spray symmetry and diffusion range. The local attention head focuses on refining the deep features of local regions of the feature map (especially high spatial weight regions), enhancing subtle features related to local anomalies (such as large single droplets, discontinuous spray, and eccentric spray). The outputs of the two heads are adaptively fused along the channel dimension to form a composite feature representation that simultaneously encodes "whether the overall system is normal" and "where the local system is abnormal." This design directly addresses the practical need in atomization state classification to comprehensively evaluate both global and local indicators.

[0073] According to embodiments of this disclosure, a training method for a fogging state recognition model is provided. The fogging state recognition model includes an asymmetric convolution module, a spatial attention module, a dual-head attention module, and a classification module. The method includes: performing multi-dimensional asymmetric convolution processing on fogging image samples using the asymmetric convolution module to obtain multi-dimensional fogging feature samples. The fogging image samples are obtained by preprocessing a fogging video of a fogging nozzle. The multi-dimensional fogging features are used to characterize the fogging region in the target fogging image from both horizontal and vertical dimensions. The method further includes performing weight masking processing on the multi-dimensional fogging feature samples based on a spatial attention mechanism using the spatial attention module to obtain a fogging feature sample with droplet boundaries. The process involves: 1) collecting spray contour feature samples containing information about the spray cone structure; 2) performing weighted fusion processing on the spray contour feature samples using a dual-head attention module to obtain atomization fusion feature samples; 3) classifying the atomization fusion feature samples using a classification module to obtain the atomization state prediction results of the atomizing nozzle; 4) processing the atomization state prediction results and the label values ​​of the atomization image samples using a preset cross-entropy loss function to obtain cross-entropy loss values, and then using these cross-entropy loss values ​​to optimize the parameters of the atomization state recognition model; and 5) iteratively performing data processing and parameter optimization operations on the atomization state recognition model until the preset training conditions are met, resulting in a trained atomization state recognition model.

[0074] The above embodiments of this disclosure relate to the training process of a fogging state recognition model. The training process provided by this disclosure will be further described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0075] Figure 3 An architecture diagram of model training and inference according to an embodiment of the present disclosure is shown.

[0076] like Figure 3As shown, the model training includes dataset construction, fogged image preprocessing, iterative model training, model inference to obtain fogged state recognition results, and alarms based on the fogged state recognition results. The model is based on an attention-enhanced asymmetric reparameterized backbone network, such as RepVGG and ACB (Asymmetric Convolutional Network). During training, the model employs multi-branch asymmetric convolutional blocks to enhance feature diversity and directional robustness, and introduces a hierarchical spatial and dual-head attention mechanism to adaptively model the spatial importance distribution and multi-scale features of fogged images. During inference, structural reparameterization technology effectively fuses multi-branch convolutions into single-branch standard convolutions, ensuring no additional computational overhead, thus achieving a balance between powerful feature modeling capabilities during training and efficient deployment performance during inference. The fogged state recognition model provided in this disclosure integrates and adaptively improves existing technologies, forming a highly specialized and inference-efficient feature extraction backbone network specifically designed to enhance the robustness and accuracy of fogged state recognition.

[0077] In this embodiment, a RepVGG_ACB fogging state recognition model that incorporates asymmetric convolutional blocks and structural reparameterization is used to process the fogged image sample dataset. During the convolutional feature extraction process of the basic RepVGG network, a model containing 3... The system employs a multi-branch structure with three convolutional branches, asymmetric convolutional branches, and identity mapping branches. After the convolutional feature output, spatial attention and dual-head attention mechanisms are introduced. The spatial attention mechanism adaptively weights the response of the feature map in the spatial dimension to enhance the feature representation of key nozzle atomization regions and suppress background interference. The dual-head attention mechanism models the global semantic features and local discriminative features of the atomized image, respectively, and fuses the output features of the two attention heads in the channel dimension to form a joint feature representation, which is then input into the classification layer. During training, the model parameters are updated in reverse using the cross-entropy loss function.

[0078] According to embodiments of this disclosure, the aforementioned atomized image samples are obtained by preprocessing the atomization video of the atomizing nozzle, including: decoding the atomization video to obtain a sequence of atomized image frames; filtering the sequence of atomized image frames according to preset filtering criteria to obtain a target atomized frame sequence, wherein the preset filtering criteria include image brightness, image contrast, and gamma distribution information of the image; and using a preset algorithm to enhance the droplet edges of the target atomized frame sequence to obtain atomized image samples with removed imaging noise, wherein the preset algorithm includes a histogram equalization algorithm and a bilateral filtering algorithm.

[0079] The above embodiments of this disclosure relate to the acquisition and preprocessing of atomized image samples. In one embodiment, for example, on a tobacco feeding production line in a cigarette manufacturing enterprise, an industrial camera (frame rate set to approximately 30-40fps) is installed near the feeding nozzle to collect atomized image data of the nozzle during a batch of production in real time. The received atomized video is decoded frame by frame to obtain clear and continuous nozzle atomized image samples. Based on image brightness, contrast, and gamma distribution characteristics, quality screening is performed on the decoded atomized images to remove overly dark, overexposed, or blurry frames, retaining only valid images suitable for identification. Preprocessing methods such as histogram equalization and bilateral filtering are applied to the valid atomized images to enhance droplet edge features and suppress noise and light fog effects caused by high temperature and humidity environments. The preprocessed images are incorporated into a training set construction module, and a dataset of six atomization states is established through manual classification or automatic annotation. Enhancement operations such as rotation, flipping, and cropping are then performed on the images.

[0080] According to embodiments of this disclosure, the above-mentioned multi-dimensional asymmetric convolution processing of fogged image samples using an asymmetric convolution module to obtain multi-dimensional fogged feature samples includes: performing asymmetric convolution processing on the fogged image samples in the horizontal dimension using the horizontal convolution units of the asymmetric convolution module to obtain horizontal dimension fogged feature samples; performing asymmetric convolution processing on the fogged image samples in the vertical dimension using the vertical convolution units of the asymmetric convolution module to obtain vertical dimension fogged feature samples; and using a structure reparameterization technique, weightedly fusing the horizontal fogged feature samples and the vertical dimension fogged feature samples using the asymmetric convolution module to obtain multi-dimensional fogged feature samples.

[0081] The following detailed description, in conjunction with specific implementation methods and accompanying drawings, provides a further detailed explanation of the multi-dimensional fogging feature extraction process based on asymmetric convolution processing provided in the embodiments of this disclosure.

[0082] Figure 4 A comparative schematic diagram of horizontal convolution and standard convolution according to embodiments of the present disclosure is shown, wherein, Figure 4 In the diagram, (a) represents the horizontal convolution operation. Figure 4 In the diagram, (b) represents the standard convolution operation.

[0083] like Figure 4 As shown in (a), horizontal convolution: the input feature map (i.e., the green grid) and the red convolution kernel (i.e., the rectangular selected area) are subjected to standard convolution calculation, and the red square in the output feature map is the convolution result; flipping: the convolution kernel is flipped horizontally (i.e., symmetrically flipped left and right) to prepare for subsequent symmetrical convolution. Figure 4As shown in (b), squared convolution: the flipped convolution kernel is convolved with the input feature map, and the red square in the output feature map is the result; flip: the convolution kernel is flipped in the vertical direction (i.e., flipped symmetrically up and down) to complete the symmetry processing of symmetric convolution.

[0084] Figure 5 The training and inference processes of an asymmetric convolution module according to an embodiment of this disclosure are illustrated.

[0085] Figure 6 The RepVGG_ACB network architecture according to an embodiment of this disclosure is shown.

[0086] Figure 7 The training and inference phases of the RepVGG_ACB network according to embodiments of this disclosure are shown, wherein... Figure 7 In this context, (a) represents the asymmetric convolutional module during the training phase; Figure 7 In the diagram, (b) represents the asymmetric convolution module during the inference phase.

[0087] Fog images captured in industrial settings may undergo unexpected directional changes (such as vertical or horizontal flipping) due to camera installation angles, post-maintenance reinstallation, or other reasons. Traditional square convolution kernels (such as 3D) 3) The extracted features are sensitive to such geometric symmetry transformations, which may lead to inconsistent feature representations of the same fogging state in images with different orientations, thus causing misjudgments. To address this issue, this disclosure specifically designs a fusion method for asymmetric convolutions for fogging state discrimination tasks. During the training phase, for the asymmetric convolution module, in each basic 3 In the 3 convolutional layers, 1 is introduced in parallel. 3 horizontal convolutions and 3 1. Vertical convolutional branch. The motivation for this design is: 1 3-convolution has a stable response to spatial structures in the vertical direction (e.g., the vertical diffusion range of a spray), while 3 1. Convolution has a stable response to horizontal structures (such as the symmetry of a spray). By training in parallel with three branches, the network is forced to learn orientation-independent, more essential fog structure features from the original image and its possible flipped variants, thereby improving the model's inherent robustness to image orientation changes in real-world applications.

[0088] Figure 5 The training and inference phases of an asymmetric convolution module are illustrated. To achieve high-precision recognition with limited industrial computing resources, this disclosure designs a reparameterizable asymmetric convolution module as the basic unit for feature extraction. During the training phase, the module is presented as containing standard 3 3.1 3, 3 1. A multi-branch structure with three convolutional branches (e.g.) Figure 6 The RepVGG_ACB network architecture is shown. This multi-branch structure greatly enriches the model's ability to capture multi-directional texture and edge information in fogged images during training. Addressing the need to extract key features such as droplet boundaries and spray contours in fogged images, multi-branch collaborative training enables the network to learn a more comprehensive spatial representation. During the inference (or deployment) phase, through structural reparameterization, the convolutional kernels and batch normalization layer parameters of the three branches are mathematically transformed and equivalently fused into a single 3D model. 3 convolution kernels (e.g.) Figure 7 (As shown). This restores the inference network to a simple single-path structure, with the exact same computation graph and latency as the original RepVGG network, achieving "training gain, lossless inference," perfectly adapting to the dual requirements of high precision and high efficiency in industrial scenarios.

[0089] Figure 8 A structural diagram of a spatial attention module and a dual-head attention module according to embodiments of the present disclosure is shown.

[0090] Figure 9 A typical atomization state diagram according to an embodiment of the present disclosure is shown; wherein Figure 9 (a) in the text indicates that the injection angle is too small; Figure 9 (b) indicates that the spray angle is too large; Figure 9 (c) in the text indicates leakage; Figure 9 In this context, (d) indicates ambiguity; Figure 9 (e) in the text indicates the normal state; Figure 9 (f) in the text indicates that no liquid was sprayed.

[0091] like Figure 8 As shown, this disclosure provides a hierarchical attention mechanism to address the spatial heterogeneity of atomized regions. Atomization state recognition faces two core challenges: first, the effective spray area has a variable proportion in the image and is often accompanied by complex background interference; second, it is necessary to simultaneously evaluate macroscopic spray morphology (such as cone angle and uniformity) and microscopic local anomalies (such as dripping, line flow, and splashing). To address these challenges, this disclosure sequentially cascades a spatial attention module and a dual-head attention module on the deep features extracted by the reparameterized backbone network, forming a hierarchical feature refinement and decision enhancement process.

[0092] (1) Spatial Attention Mechanism: Focusing on Key Spray Regions. This module receives the feature map from the backbone network and learns to generate a two-dimensional spatial weight mask. This mask can highlight the spatial regions most relevant to the atomization state (such as the dense spray core area near the nozzle exit and the cone-shaped area of ​​stable spray), while suppressing the response of irrelevant backgrounds (such as equipment supports and background panels). This allows the subsequent computational resources of the network to be concentrated on the image regions that are truly useful for discrimination, directly addressing the problem of the non-fixed spatial distribution of target regions in atomized images.

[0093] (2) Dual-head attention mechanism: collaborative modeling of macroscopic morphology and microscopic anomalies. Based on the features after spatial attention filtering, this disclosure designs a dual-head attention mechanism specifically tailored for the fogging state. This mechanism includes two parallel attention branches with different functional focuses:

[0094] Global Attention Head: Focuses on modeling the channel relationships of the entire feature map to capture and enhance feature channels related to the overall spray pattern, such as the overall pattern reflecting spray symmetry and diffusion range.

[0095] Local attention head: This head focuses on refining the deep features of local regions of the feature map (especially regions with high spatial weights) to enhance subtle features related to local anomalies (such as single large droplets, discontinuous spray, and eccentric spray). The outputs of the two heads are adaptively fused along the channel dimension to form a composite feature representation that simultaneously encodes "whether the overall system is normal" and "where the local system is abnormal." This design directly addresses the practical need in atomization state classification to comprehensively evaluate both global and local indicators (six typical states, such as...). Figure 9 As shown in the figure, it significantly improves the model's ability to distinguish between "normal uniform spray" and "various abnormal sprays".

[0096] This disclosure introduces the RepVGG_ACB series network with structural reparameterization into the task of identifying the atomization state of the feeding nozzle, and further combines spatial attention mechanism and dual-head attention mechanism to achieve adaptive focusing on key atomization areas and collaborative modeling of global and local features, which significantly improves the overall performance of the identification system in complex industrial scenarios, specifically in the following aspects.

[0097] (1) Significantly improve recognition accuracy and achieve stronger feature representation capability: By introducing a spatial attention mechanism into the RepVGG_ACB network structure, the model can automatically focus on key atomization areas such as spray cone structure and droplet dense area during feature extraction, effectively suppressing interference from complex background and irrelevant areas; at the same time, by using a dual-head attention mechanism to collaboratively model global spray morphology features and local abnormal detail features, the model's ability to distinguish different atomization states is further enhanced, thereby significantly improving recognition accuracy and feature representation capability.

[0098] (2) While ensuring high accuracy, the reasoning speed is greatly improved and the recognition latency is reduced: This disclosure introduces spatial attention mechanism and dual-head attention mechanism, and combines structural reparameterization technology to compress the multi-branch structure into a single-branch convolutional structure in the reasoning stage. This effectively avoids the negative impact of attention mechanism on reasoning efficiency, achieves a good balance between high accuracy and fast reasoning, and meets the needs of real-time detection in industrial sites.

[0099] (3) Significantly reduce model size and computational overhead, and improve system deployment efficiency: Compared with traditional large-scale convolutional networks, the RepVGG_ACB series models disclosed in this paper maintain a small model size and computational overhead while introducing an attention mechanism, which is convenient for deployment in resource-constrained industrial environments, without affecting recognition performance.

[0100] (4) The accuracy is better than that of the lightweight model, while the speed remains leading, achieving a better balance between accuracy and efficiency: Compared with the existing lightweight network, this disclosure introduces spatial attention mechanism and dual-head attention mechanism, which enables the model to significantly improve the accuracy of fogging state recognition while maintaining high inference speed, achieving a better balance between accuracy and efficiency under limited computing resources, and enhancing the industrial applicability of the model.

[0101] (5) The synergistic effect of structural reparameterization and attention mechanism has excellent engineering adaptability: During the training phase, the RepVGG_ACB network enhances the feature representation ability through the synergistic effect of multi-branch structure, spatial attention mechanism and dual-head attention mechanism; during the inference phase, the multi-branch structure is compressed into a single-branch structure through structural reparameterization, avoiding the adverse effect of attention mechanism on inference efficiency.

[0102] This design enables the recognition model used in this disclosure to have engineering advantages such as simple structure, stable operation, and easy integration in actual industrial deployments, and can simultaneously meet the comprehensive requirements of industrial sites for real-time performance, recognition accuracy, system stability, and computing cost.

[0103] Figure 10 A structural block diagram of an atomization state recognition device for an atomizing nozzle according to an embodiment of the present disclosure is shown.

[0104] like Figure 10 As shown, the atomization state recognition device 1000 of the atomizing nozzle in this embodiment includes an imaging noise removal module 1010, an asymmetric convolution module 1020, an attention mechanism module 1030, and an atomization state classification module 1040.

[0105] The imaging noise removal module 1010 is used to perform droplet edge enhancement processing on the original atomized image generated by the atomizing nozzle during the production process to obtain a target atomized image with imaging noise removed. The imaging noise includes imaging noise generated by the atomizing nozzle due to the light fog effect and imaging noise generated by water vapor particles during the production process. In one embodiment, the imaging noise removal module 1010 can be used to perform the operation S210 described above, which will not be repeated here.

[0106] The asymmetric convolution module 1020 is used to perform multi-dimensional asymmetric convolution processing on the target fogged image to obtain multi-dimensional fogging features. The multi-dimensional fogging features are used to characterize the features of the fogged region in the target fogged image from the horizontal and vertical dimensions. In one embodiment, the asymmetric convolution module 1020 can be used to perform the operation S220 described above, which will not be repeated here.

[0107] The attention mechanism module 1030 is used to perform weight masking processing on multi-dimensional atomization features based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information, and to perform weight fusion processing on spray contour features based on dual-head attention mechanism to obtain atomization fusion features; in one embodiment, the attention mechanism module 1030 can be used to perform the operation S230 described above, which will not be repeated here.

[0108] The atomization state classification module 1040 is used to classify the atomization fusion features to obtain the atomization state identification result of the atomizing nozzle. In one embodiment, the atomization state classification module 1040 can be used to perform the operation S240 described above, which will not be repeated here.

[0109] According to embodiments of this disclosure, any multiple modules among the imaging noise removal module 1010, asymmetric convolution module 1020, attention mechanism module 1030, and fog state classification module 1040 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the imaging noise removal module 1010, asymmetric convolution module 1020, attention mechanism module 1030, and fog state classification module 1040 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the imaging noise removal module 1010, the asymmetric convolution module 1020, the attention mechanism module 1030, and the fogging state classification module 1040 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0110] Figure 11 A block diagram of an electronic device suitable for implementing a fogging state recognition method according to an embodiment of the present disclosure is shown.

[0111] like Figure 11 As shown, an electronic device 1100 according to an embodiment of the present disclosure includes a processor 1101, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1102 or a program loaded from a storage portion 1108 into a random access memory (RAM) 1103. The processor 1101 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1101 may also include onboard memory for caching purposes. The processor 1101 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0112] RAM 1103 stores various programs and data required for the operation of electronic device 1100. Processor 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Processor 1101 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 1102 and / or RAM 1103. It should be noted that the programs may also be stored in one or more memories other than ROM 1102 and RAM 1103. Processor 1101 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0113] According to embodiments of this disclosure, the electronic device 1100 may further include an input / output (I / O) interface 1105, which is also connected to a bus 1104. The electronic device 1100 may also include one or more of the following components connected to the input / output (I / O) interface 1105: an input section 1106 including a keyboard, mouse, etc.; an output section 1107 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a LAN card, modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the input / output (I / O) interface 1105 as needed. A removable medium 1111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1110 as needed so that computer programs read from it can be installed into the storage section 1108 as needed.

[0114] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0115] According to embodiments of this disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this disclosure, the computer-readable storage medium may include ROM 1102 and / or RAM 1103 and / or one or more memories other than ROM 1102 and RAM 1103 described above.

[0116] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this disclosure.

[0117] When the computer program is executed by the processor 1101, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0118] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1109, and / or installed from the removable medium 1111. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0119] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 1109, and / or installed from removable medium 1111. When the computer program is executed by processor 1101, it performs the functions defined in the system of this disclosure embodiment. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0120] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0121] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0122] Those skilled in the art will understand that the features described in the various embodiments of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments of this disclosure can be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0123] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for identifying the atomization state of an atomizing nozzle, characterized in that, include: The original atomized image generated by the atomizing nozzle during the production process is subjected to droplet edge enhancement processing to obtain a target atomized image with imaging noise removed. The imaging noise includes imaging noise generated by the atomizing nozzle during the production process due to the light fog effect and imaging noise generated by water vapor particles. The target fogged image is subjected to multi-dimensional asymmetric convolution processing to obtain multi-dimensional fogging features, wherein the multi-dimensional fogging features are used to characterize the features of the fogged region in the target fogged image from the horizontal and vertical dimensions. The multi-dimensional atomization features are subjected to weight masking based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information. The spray contour features are then subjected to weight fusion processing based on dual-head attention mechanism to obtain atomization fusion features. The atomization fusion features are classified to obtain the atomization state identification result of the atomizing nozzle.

2. The method according to claim 1, characterized in that, The original atomized image generated during the production process of the atomizing nozzle is subjected to droplet edge enhancement processing to obtain a target atomized image with imaging noise removed, including: The grayscale histogram of the original fogged image is obtained by counting the number of pixels at each gray level in the original fogged image; The gray-level cumulative distribution function is calculated using the gray-level histogram, and the gray-level values ​​in the original fogged image are remapped using the gray-level cumulative distribution function to enhance the fog droplet edges in the original fogged image, thereby obtaining the target fogged image with imaging noise removed.

3. The method according to claim 1, characterized in that, The original atomized image generated during the production process of the atomizing nozzle is subjected to droplet edge enhancement processing to obtain a target atomized image with imaging noise removed, including: Calculate the spatial distance between pixels in the original fogged image to obtain the spatial domain weight, and calculate the similarity between pixel values ​​in the original fogged image to obtain the pixel value domain weight; The spatial domain weights and the pixel value domain weights are calculated to obtain the fusion weights; The original fogged image is processed by pixel-weighted averaging using the fusion weights to enhance the edges of fog droplets in the original fogged image, thereby obtaining the target fogged image with imaging noise removed.

4. The method according to claim 1, characterized in that, The target fogged image is subjected to multi-dimensional asymmetric convolution processing to obtain multi-dimensional fogging features, including: The target fogged image is subjected to asymmetric convolution processing in the horizontal dimension to obtain horizontal fogging features, and the target fogged image is subjected to asymmetric convolution processing in the vertical dimension to obtain vertical fogging features. Based on structural reparameterization technology, the horizontal atomization features and the vertical dimension atomization features are weighted and fused to obtain the multi-dimensional atomization features.

5. The method according to claim 1, characterized in that, The multi-dimensional atomization features are subjected to weight masking based on a spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information, including: The multi-dimensional fogging features are subjected to global average pooling to obtain the multi-dimensional fogging features after global average pooling, and the multi-dimensional fogging features are subjected to global max pooling to obtain the multi-dimensional fogging features after global max pooling. The multi-dimensional fogging features after global average pooling and the multi-dimensional fogging features after global max pooling are subjected to convolutional activation processing in the channel dimension to obtain a spatial attention weight map. The spray contour features are obtained by weighting the spatial attention weight map and the multi-dimensional atomization features to enhance the droplet boundaries and spray cone structure of the atomization region.

6. The method according to claim 5, characterized in that, The spray contour features are subjected to weighted fusion processing based on a dual-head attention mechanism to obtain atomization fusion features, including: The spray profile features are enhanced in the channel dimension using a global attention head mechanism to obtain global spray features, wherein the spray pattern includes the symmetry of the spray and the diffusion range of the spray. The spray profile features are enhanced with local spray characteristics based on a local attention head mechanism to obtain local spray characteristics, wherein the local spray characteristics include droplet shape, spray angle and spray splash state; The global spray state and the local spray state are fused based on the weighted weights of the spatial attention weight map to obtain the atomization fusion feature.

7. A device for identifying the atomization state of an atomizing nozzle, characterized in that, include: An imaging noise removal module is used to perform droplet edge enhancement processing on the original atomized image generated by the atomizing nozzle during the production process to obtain a target atomized image with imaging noise removed. The imaging noise includes imaging noise generated by the atomizing nozzle due to the light fog effect and imaging noise generated by water vapor particles during the production process. An asymmetric convolution module is used to perform multi-dimensional asymmetric convolution processing on the target fogged image to obtain multi-dimensional fogging features, wherein the multi-dimensional fogging features are used to characterize the features of the fogged region in the target fogged image from the horizontal and vertical dimensions. The attention mechanism module is used to perform weight masking processing on the multi-dimensional atomization features based on spatial attention mechanism to obtain spray contour features with droplet boundary information and spray cone structure information, and to perform weight fusion processing on the spray contour features based on dual-head attention mechanism to obtain atomization fusion features. The atomization state classification module is used to classify the atomization fusion features to obtain the atomization state identification result of the atomizing nozzle.

8. A training method for a fogging state recognition model, wherein the fogging state recognition model includes an asymmetric convolution module, a spatial attention module, a dual-head attention module, and a classification module, characterized in that, include: The asymmetric convolution module is used to perform multi-dimensional asymmetric convolution processing on the fogged image samples to obtain multi-dimensional fogging feature samples. The fogged image samples are obtained by preprocessing the fogging video of the fogging nozzle. The multi-dimensional fogging features are used to characterize the features of the fogging region in the target fogged image from the horizontal and vertical dimensions. The spatial attention module is used to perform weight masking processing on the multi-dimensional atomization feature samples based on the spatial attention mechanism to obtain spray contour feature samples with droplet boundary information and spray cone structure information; The dual-head attention module is used to perform weight fusion processing on the spray contour feature samples based on the dual-head attention mechanism to obtain atomization fusion feature samples; The classification module is used to classify the atomization fusion feature samples to obtain the atomization state prediction result of the atomization nozzle; The fogging state prediction result and the label value of the fogging image sample are processed using a preset cross-entropy loss function to obtain the cross-entropy loss value, and the parameters of the fogging state recognition model are optimized using the cross-entropy loss value. The data processing and parameter optimization operations of the fogging state recognition model are performed iteratively until the preset training conditions are met, and the trained fogging state recognition model is obtained.

9. The method according to claim 8, characterized in that, The atomized image samples are obtained by preprocessing the atomization video of the atomizing nozzle, including: The fogged video is decoded to obtain a fogged image frame sequence; The fogged image frame sequence is filtered according to a preset filtering standard to obtain a target fogged frame sequence, wherein the preset filtering standard includes image brightness, image contrast and image gamma distribution information; The target fogged frame sequence is subjected to fog droplet edge enhancement processing using a preset algorithm to obtain fogged image samples with imaging noise removed. The preset algorithm includes a histogram equalization algorithm and a bilateral filtering algorithm.

10. The method according to claim 8, characterized in that, The asymmetric convolution module is used to perform multi-dimensional asymmetric convolution processing on the fogged image samples to obtain multi-dimensional fogged feature samples, including: The horizontal convolution unit of the asymmetric convolution module is used to perform asymmetric convolution processing on the fogged image sample in the horizontal dimension to obtain a fogged feature sample in the horizontal dimension. The vertical convolution unit of the asymmetric convolution module is used to perform asymmetric convolution processing on the fogged image samples in the vertical dimension to obtain vertical dimension fogged feature samples. Based on the structural reparameterization technique, the horizontal fogging feature samples and the vertical dimension fogging feature samples are weighted and fused using the asymmetric convolution module to obtain the multi-dimensional fogging feature samples.

Images