Intelligent operation system of air force road marking line removing machine based on AI image recognition

Abstract

The application provides an AI image recognition-based pneumatic road marking line removing machine intelligent operation system, and relates to the technical field of computer vision, and comprises: an image acquisition module, which is used for acquiring original images containing road marking lines and corresponding ambient light intensity values of each image; an image processing module, which is used for sequentially performing contrast enhancement and deblurring processing on the original images; an image recognition module, which is used for analyzing and recognizing the processed images by adopting a double-attention-guided consistency network, and outputting type, wear degree and covered area coordinate information of the road marking lines; a parameter decision module, which is used for adaptively matching and adjusting line removing pressure and operation speed parameters of the pneumatic line removing machine according to the result output by the image recognition module; and an execution control module, which is used for controlling an execution mechanism of the line removing machine to realize closed-loop control of line removing parameters. The application can efficiently, accurately and adaptively perform line removing operation, effectively avoid road surface damage and significantly reduce manual intervention cost.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an intelligent operating system for a pneumatic road marking removal machine based on AI image recognition. Background Technology

[0002] Pneumatic road marking removal machines rely on high-pressure airflow to impact and remove road markings. Their effectiveness is highly dependent on the matching degree between parameters such as removal pressure and operating speed and the actual condition of the markings. Traditional removal machines require manual observation of the marking type, wear level, and surrounding environment, and manual parameter adjustments, resulting in problems such as low recognition accuracy, delayed parameter adjustments, low operating efficiency, and potential damage to the road surface.

[0003] While existing AI image recognition technologies have made breakthroughs in target recognition, image deblurring, and data enhancement, they still face multi-dimensional technical bottlenecks in pneumatic line removal machine scenarios: First, road images are easily blurred due to changes in lighting and dust interference, resulting in blurred edge features of road markings. Existing deblurring methods mostly adopt a uniform processing strategy, which cannot adaptively identify the degree of blurring in different areas, making it difficult to accurately preserve the details of the road markings. Second, the grayscale difference between road markings and the road surface is small, especially in scenarios with worn road markings, resulting in low feature contrast. Traditional image enhancement techniques (such as simple Gamma correction and histogram equalization) cannot effectively address these issues. First, existing target recognition networks often employ a single attention mechanism or a fixed decoder structure, which is insufficient in distinguishing road marking types (such as solid lines and dashed lines) and determining wear levels (light / moderate / severe). Furthermore, the models have weak generalization ability under limited samples and are easily affected by road surface stains, shadows, etc. Second, during model training, the confidence differences of the probability maps output by a single decoder are not significant, and there is a lack of effective consistency constraint mechanisms, resulting in poor stability of prediction results and difficulty in reliably guiding the adjustment of line parameters.

[0004] Specifically, current technologies have significant shortcomings in the following aspects: Image enhancement: Traditional grayscale enhancement techniques do not incorporate the lighting and reflection characteristics of road images for layered processing, failing to specifically enhance the grayscale differences between road markings and the road surface. Especially in low-light or high-light scenarios, the enhancement effect is limited, making it difficult to meet the high-contrast feature requirements of subsequent recognition. Image deblurring: Existing deblurring networks lack a dedicated module for blur region recognition, lacking specificity in handling different types of blur such as dust and motion blur. This easily leads to overly smoothed or residual blur at the edges of road markings, affecting feature extraction quality. Feature processing: Most networks do not combine feature difference operations with channel attention mechanisms, failing to effectively highlight key features in blurry areas while suppressing redundant information from the background road surface, resulting in insufficient feature representation capabilities. Target recognition: Existing recognition networks mostly employ a single decoder structure, lacking differentiated capture of different road marking features and failing to introduce a dual attention mechanism (spatial + channel) to focus on road marking areas, resulting in low recognition accuracy for worn markings and fine dashed lines. At the model training level: there is a lack of effective pseudo-label generation and consistency constraint strategies, the sharpening effect of a single probability map is poor, the prediction consistency between different decoders is weak, and the model's generalization ability and stability are difficult to meet the needs of complex road scenarios.

[0005] Current technology cannot meet the high-efficiency and precise requirements of pneumatic road marking removal machines. Summary of the Invention

[0006] To address the aforementioned problems, this invention aims to solve the issues of traditional pneumatic road marking removal machines relying on manual labor, resulting in low efficiency and poor accuracy. Simultaneously, it overcomes the multi-dimensional bottlenecks of existing AI image recognition technology in road marking removal scenarios. This invention integrates multi-module AI technologies (layered enhancement, adaptive deblurring, dual attention recognition, and consistency training) into the innovative intelligent operation technology of the road marking removal machine, achieving intelligent identification of road marking conditions and automatic adjustment of removal parameters, thereby improving the intelligence level, efficiency, and accuracy of the removal operation.

[0007] To achieve the above objectives, the following technical solution is adopted:

[0008] This invention provides an intelligent operating system for a pneumatic road marking removal machine based on AI image recognition. The system includes:

[0009] The image acquisition module is used to acquire original images containing road markings and the ambient light intensity value for each image.

[0010] The image processing module is used to sequentially perform contrast enhancement and deblurring on the original image to improve the clarity and distinguishability of the marking features, thereby obtaining the processed image.

[0011] The image recognition module uses a dual attention-guided consistency network (DACNet) with an architecture that includes a shared encoder and multiple differential decoders to analyze and recognize the processed images. Each differential decoder integrates a spatial attention module and a channel attention module, and optimizes the model through a complementary consistency training strategy to output the type, wear level and coverage area coordinates of road markings.

[0012] The parameter decision module is used to adaptively match and adjust the wire removal pressure and operating speed parameters of the pneumatic wire removal machine based on the results output by the image recognition module.

[0013] The execution control module is used to control the actuator of the wire removal machine according to the output signal of the parameter decision module, including a pressure regulating valve and a walking drive motor, to realize closed-loop control of the wire removal parameters.

[0014] Compared with the prior art, the present invention achieves the following beneficial effects:

[0015] 1. This invention employs grayscale layering technology: addressing the issues of uneven illumination in road images and small grayscale differences between road markings and the road surface, the image is decomposed into illumination and reflection images through logarithmic transformation. Combined with vector grid layering processing and three-dimensional Gaussian filtering, the feature contrast of the road image is enhanced, providing a high-quality feature foundation for subsequent deblurring and recognition, and improving the distinguishability of road marking features.

[0016] 2. This invention proposes an enhancement scheme that combines histogram truncation with improved Gamma correction: To address the problem that traditional enhancement techniques tend to over-enhance noise or lose details, the histogram truncation method is used to extract abnormal pixels at both ends of the illuminated image. Combined with an improved Gamma correction formula (introducing brightness gain parameters and contrast control parameters), adaptive enhancement for different lighting scenarios (weak light, strong light) is achieved, optimizing the grayscale difference between road markings and the road surface, while suppressing noise and ensuring the integrity of details in the enhanced image.

[0017] 3. This invention designs an adaptive blurred region recognition network (AIBNet): To address the quality issues of road images caused by dust and motion blur, AIBNet, which includes a multi-sub-decoder, is designed. Through the spatial feature difference processing block (SFDHBlock) and the high-frequency feature selection block (HFSBlock), it adaptively identifies blurred regions in the image, specifically eliminates different types of blur, preserves the edge details of the road markings, and improves the quality of the deblurred image, laying the foundation for accurate recognition.

[0018] 4. The present invention designs a spatial feature difference processing block (SFDHBlock): To address the problem of difficulty in highlighting features in blurred regions, an SFDHBlock is designed, consisting of a spatial feature enhancement module (SFEM) and simple channel attention (SCA). SFEM amplifies the feature differences between blurred and clear regions through feature difference operations (difference between two sets of attention weight maps), while SCA strengthens the effective features of the channel dimension. After weighted fusion, the two significantly improve the feature representation ability of blurred regions while suppressing background interference.

[0019] 5. This invention proposes a Dual Attention Guided Consistency Network (DACNet): To address the problem of insufficient accuracy in recognizing marking types and wear levels, a DACNet is constructed that includes one shared encoder and four differentiated decoders. In the decoders, a spatial attention module (focusing on the spatial position of the marking) and a channel attention module (enhancing the channel features related to the marking) are introduced. By adjusting the jump connection structure, the decoder is differentiated, thereby improving the recognition accuracy of different types of markings and different wear levels, and outputting reliable marking type, wear level, and coverage area coordinates.

[0020] 6. This invention proposes a complementary consistency training strategy: to address the problems of low confidence in probability maps and poor prediction stability during model training, the probability maps output by each decoder are sharpened (by introducing a temperature parameter to control the sharpening intensity) to generate high-confidence pseudo-labels. The consistency loss between different decoders is calculated using mean squared error (MSE), thereby enhancing the consistency and stability of model predictions, improving generalization ability under limited samples, and ensuring that the recognition results reliably guide the adjustment of line parameters.

[0021] In summary, this invention integrates the above-mentioned technological innovations to achieve intelligent processing of the entire process from road image acquisition, enhancement, deblurring, recognition to parameter decision-making. This enables the pneumatic line removal machine to adapt to different road scenarios and automatically match the optimal line removal parameters, thereby improving line removal efficiency, reducing road surface damage, and lowering the cost of manual intervention.

[0022] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0023] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0024] Figure 1This is a schematic diagram of a module of an intelligent operating system for a pneumatic road marking removal machine based on AI image recognition, according to an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the intelligent operation system of a pneumatic road marking removal machine based on AI image recognition, according to an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the structure of the Dual Attention Guided Consistency Network (DACNet) in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0029] Figure 1 This is a schematic diagram of a module of an intelligent operating system for a pneumatic road marking removal machine based on AI image recognition, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the operation of an intelligent pneumatic road marking removal machine system based on AI image recognition, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the present invention discloses an intelligent operating system 100 for a pneumatic road marking removal machine based on AI image recognition. The system 100 includes:

[0030] Image acquisition module 110 is used to acquire original images containing road markings and ambient light intensity values ​​for each image.

[0031] In some embodiments of the present invention, the image acquisition module 110 is used for image acquisition. Specifically, a high-definition industrial camera (such as Hikvision MV-CA050-10GM) with a resolution of not less than 1920×1080 and a frame rate of not less than 30fps is used. Two cameras are symmetrically installed at the front end of the road marking removal machine, with the acquisition angle covering the road marking removal operation area (3-5m wide, 0.5-2m away from the camera), to acquire the original image containing the road markings. Simultaneously, the camera's built-in light intensity sensor is activated to record the ambient light intensity value (unit: lux) at the time of each image acquisition, and stores it as lighting-assisted data.

[0032] In some optional embodiments of the present invention, the system further includes a lidar module, which synchronously acquires data with the image acquisition module 110, specifically acquiring three-dimensional point cloud data and reflectivity information of the road surface as auxiliary data for cross-validation or supplementary judgment of the road marking recognition results in complex scenarios. Specifically, the system is equipped with a lidar (such as the RoboSense RS-LIDAR-M1, with an acquisition range of 0.1-200m and a point cloud density of...). Simultaneously collect three-dimensional point cloud data of the road surface to obtain road height and reflectivity information. Among them, reflectivity information can directly reflect the reflection characteristics of the road surface (markings and road surface) to laser; three-dimensional height information is used to eliminate the influence of terrain interference such as road surface protrusions and depressions on the light judgment.

[0033] The image processing module 120 is used to sequentially perform contrast enhancement and deblurring processing on the original image to improve the clarity and distinguishability of the marking features, thereby obtaining the processed image.

[0034] (1) Enhance the contrast of road images

[0035] This invention proposes a grayscale layering technique to enhance the feature contrast of road images and improve subsequent recognition accuracy. To this end, the image processing module 120 includes an image enhancement unit 121 for enhancing the contrast of the original image. The enhancement process includes: decomposing the image using grayscale layering technology, and performing enhancement operations combining histogram truncation and Gamma correction on the decomposed illumination components to obtain the enhanced road image; wherein, the parameters of the Gamma correction are dynamically configured according to the ambient light intensity value, adaptively adjusting the enhancement processing strategy.

[0036] The grayscale layering technique specifically includes: performing a logarithmic transformation on the original image to decompose it into an illumination image and a reflection image; constructing a vector grid corresponding to the original image and smoothing the vector grid using a three-dimensional Gaussian filter; and further optimizing the filtered vector grid using a bilateral filter to preserve image edge details. The specific steps are as follows.

[0037] Step S1.1: Preprocessing of the original image

[0038] The original road image acquired by the image acquisition module 110 Perform a logarithmic transformation to change the illumination image in the image. With reflected image The product relationship can be converted into a summation relationship using the following formula:

[0039] ln: Natural logarithm function, used to transform the product relationship between the illumination image and the reflection image in the original image into a summation relationship, reducing the difficulty of feature extraction caused by differences in image brightness. Original road image in coordinates Pixel value at that location, The coordinates of the pixel in the horizontal direction (horizontal axis) of the image. The coordinates of the pixel in the vertical direction (vertical axis) of the image, with a value range of 0-255 (for 8-bit grayscale images). : Illumination image in coordinates The pixel value at that location reflects the illumination distribution in the original road image, and its value ranges from 0 to 255. The reflected image in coordinates The pixel value at a given location reflects the light reflection characteristics of the road surface (including markings), and the value ranges from 0 to 255.

[0040] Step S1.2: Construct a vector grid

[0041] Initialize the vector grid corresponding to the image The formula is:

[0042]

[0043] : The 3D vector grid corresponding to the original road image, with dimensions (image width) Image height (Gray level dimension), used for subsequent gray level layering processing, storing image gray level information and labeling information. : The coordinates of the vector grid in the horizontal direction (x-axis), which correspond to the pixel coordinates in the horizontal direction of the original road image, and the value range is 0-(image width-1). : The coordinates of the vector grid in the vertical direction (y-axis), corresponding to the pixel coordinates in the vertical direction of the original road image, with a value range of 0 to (image height - 1). : Gray-level dimension coordinates, whose value range is related to the image gray-level values. Consistent, ranging from 0 to 255 (for 8-bit grayscale images), this dimension is used to layer and index image information in the grayscale value space. : A two-dimensional grayscale image of a road in coordinates The grayscale value at the location is obtained by converting the original color road image through a grayscale conversion algorithm (such as the weighted average method), and the value range is 0-255. When satisfied The time vector grid takes two values: the first element is the gray value at the corresponding position, and the second element "1" is a marker value used to indicate that the grid position belongs to the effective gray information area. When not satisfied The time vector grid has three elements, all of which are 0, used to indicate that the grid position belongs to an invalid or unfilled area.

[0044] Step S1.3: Vector grid filtering

[0045] A three-dimensional Gaussian filter is applied to each layer of the vector mesh, using the following formula:

[0046]

[0047] The output vector grid after 3D Gaussian filtering, with dimensions equal to the input vector grid. Consistency is used to smooth image noise and preserve key grayscale features. The standard deviation of the Gaussian filter is used to control the decay rate of the Gaussian function, and its value ranges from 0.5 to 2.0. The smaller the value, the more concentrated the filter kernel, resulting in more image details being preserved and weaker noise suppression. The larger the value, the more dispersed the filter kernel, resulting in a stronger smoothing effect on the image but also greater loss of detail. The convolution operator represents applying a three-dimensional Gaussian filter kernel. With input vector grid By performing convolution operations and calculating the weighted average value at each grid position using a sliding filter kernel, image smoothing is achieved. The input 3D vector grid, i.e. the original road image grid data after initialization, provides the original input information for the convolution operation. : A three-dimensional Gaussian function, used to generate a Gaussian filter kernel, with the following formula:

[0048]

[0049] These are the coordinates of the filter kernel center. It is the center coordinate of the filter kernel in the horizontal direction (x-axis) of the image; It is the center coordinate of the filter kernel in the vertical direction (y-axis) of the image; It is the center coordinate of the filter kernel in the gray level dimension (t-axis).

[0050] It should be noted that the three-dimensional Gaussian filtering here is performed in a spatial domain. The filtering of the grayscale domain (t) is performed in a three-dimensional space together with the grayscale domain (t). The physical meaning of filtering the grayscale domain t is to smooth the information on adjacent grayscale levels, rather than smoothing in time or space.

[0051] Subsequently, a vector grid based on the output of a three-dimensional Gaussian filter. The reflection image is estimated through inverse transform and feature extraction. Specifically, for Reconstruction is performed in the grayscale dimension to obtain the enhanced reflection component estimate. .

[0052] Calculate the bilateral filtering result The formula is:

[0053]

[0054] The output of bilateral filtering is the pixel value of the road image in the current calculation area after grayscale layering and bilateral filtering. It is used to preserve image edge details while suppressing noise. The summation operator represents the summation of a set of pixels. All neighboring pixels The corresponding items are summed up. The set of pixels corresponding to the road image, specifically referring to the current pixel. neighboring pixels The size of the neighborhood of the set can be set according to actual needs (e.g.) (pixel region). : Neighboring pixels The corresponding reflected image pixel values ​​are obtained by decomposing the original road image, reflecting the neighborhood pixels. The surface reflection characteristics at the location. exp: an exponential function, used here to calculate Gaussian weights. It assigns different weights to pixels at different distances or with different grayscale levels based on exponential decay characteristics. The formula is... ( (This is a natural constant, approximately equal to 2.718). : Current pixel With neighboring pixels The Euclidean distance between them is used to measure the difference in their spatial location within the image. : Current pixel With neighboring pixels The square of the Euclidean distance is used to simplify the calculation of the Gaussian function and avoid the complexity caused by the square root operation. : Standard deviation of Gaussian function in spatial domain, used to control the impact of spatial location differences on weights; The smaller the pixel size, the greater the weight of pixels that are spatially closer, and the better the preservation of local image details. The larger the value, the slower the spatial location weight decays, and the stronger the image smoothing effect. The square of the standard deviation of the Gaussian function in the spatial domain is the core parameter in the calculation of Gaussian weights, used to adjust the decay rate of the weights due to spatial distance. : Current pixel The grayscale value ranges from 1 to 1. This reflects the brightness level of the pixel. : Neighboring pixels The grayscale value ranges from 1 to 1. ,and Together they are used to measure the brightness difference between two pixels. : Current pixel With neighboring pixels The absolute value of the difference in grayscale values, i.e. This is used to measure the difference between the two in the brightness domain. : Current pixel With neighboring pixels The square of the absolute value of the difference in grayscale values ​​is used to simplify the calculation of Gaussian weights in the brightness domain. : Standard deviation of Gaussian function in the brightness domain, used to control the impact of brightness differences on weights; The smaller the value, the greater the weight of pixels with similar brightness, which can better preserve the image edges (where the brightness difference is large, the weight is small, thus avoiding smoothing). The larger the value, the smaller the impact of brightness differences on weights, and the higher the overall smoothness of the image. The square of the standard deviation of the Gaussian function in the luminance domain is the core parameter for calculating the Gaussian weight in the luminance domain, which adjusts the decay rate of the weight due to luminance differences.

[0055] Step S1.4: Image enhancement operation

[0056] The image is enhanced by extracting pixels from both ends of the illuminated image using the histogram truncation method and combining this with a modified Gamma correction formula. The formula is as follows:

[0057]

[0058] : Improved Gamma correction enhances the output value at the i-th pixel position, with a value range of 0-255, used to optimize image contrast and brightness, highlighting the grayscale difference between road markings and the road surface. i: Pixel index, used to identify the position of different pixels in the image, with a value range of 0-(total number of pixels in the image - 1), corresponding to the number of image pixels. : Correction parameter, used to adjust the overall brightness gain of the enhanced image, with a value range of . ; At this time, the overall image brightness is improved, making it suitable for road scenes with low lighting; At this time, the overall brightness of the image decreases, making it suitable for overexposed road scenes under strong light; At this time, only the contrast is adjusted by the Gamma coefficient, without changing the overall brightness. Before Gamma correction, the pixel value at the i-th pixel position in the original image (or the image after preprocessing) has a value range of... It is the input basis for correction calculation. The control parameter, namely the Gamma coefficient, is used to adjust image contrast, and its value range is... ; At this time, the brightness and contrast of the dark areas of the image are increased, which is suitable for scenarios where the gray difference between the markings and the road surface is small; When the image is bright, the brightness of the bright areas is reduced and the details of the dark areas are compressed, which is suitable for scenes where the image is too bright or has severe reflections. At this time, the image contrast remains unchanged, and only the brightness gain is adjusted. The d-th power operation of the original pixel value y is the core calculation step of Gamma correction. It adjusts the pixel value distribution through non-linear transformation to achieve contrast optimization.

[0059] In some embodiments of the present invention, the parameters of Gamma correction include a brightness gain parameter a and a contrast control parameter d; the parameters of Gamma correction are dynamically configured according to the ambient light intensity value, and the enhancement processing strategy is adaptively adjusted: different value combinations are assigned to the brightness gain parameter a and the contrast control parameter d according to the preset range in which the ambient light intensity value is located, specifically including:

[0060] When the ambient light intensity value is less than or equal to the first preset threshold, the image enhancement unit 121 is configured to use a parameter combination of a>1 and d<1 to improve the overall brightness of the image and enhance the contrast of dark areas.

[0061] When the ambient light intensity value is greater than or equal to the second preset threshold, the image enhancement unit 121 is configured to use a parameter combination of a<1 and d>1 to reduce the overall brightness of the image and restore the details of the bright parts.

[0062] The first preset threshold is less than the second preset threshold. Specifically, the first and second preset thresholds are determined based on a large amount of experimental data from road scenarios. In a specific embodiment, the first preset threshold can be set to 100 lux to distinguish low-light environments; the second preset threshold can be set to 10000 lux to distinguish strong-light environments. When the ambient light intensity value is between the first and second preset thresholds, the system is configured to use a parameter combination of a=1 and d=1, that is, to maintain the original contrast and brightness of the image. The specific values ​​of parameters a and d can be obtained by querying a preset light intensity-parameter mapping table. For example, when the light intensity is ≤100 lux, the mapping is (a=1.2, d=0.6); when the light intensity is ≥10000 lux, the mapping is (a=0.9, d=1.4). It should be noted that the specific values ​​of the first and second preset thresholds (e.g., 100 lux and 10000 lux) and the corresponding parameter combinations (a, d) mentioned above can be determined by those skilled in the art through a limited number of calibration experiments based on typical road lighting conditions under different seasons and weather conditions. For example, image samples are collected at dusk, night, or inside tunnels, and their light intensity distribution is statistically analyzed to determine the first preset threshold; samples are collected under direct midday sunlight to determine the second preset threshold. The light intensity-parameter mapping table is established after testing the impact of different parameter combinations on image enhancement effects in these typical scenarios.

[0063] (2) Deblurring of road images

[0064] This invention proposes an Adaptive Blur Region Recognition Network (AIBNet) to eliminate quality problems in road images caused by dust and motion blur. To this end, the image processing module 120 further includes an image deblurring unit 122, implemented using the Adaptive Blur Region Recognition Network (AIBNet), used to deblur the enhanced road image. The Adaptive Blur Region Recognition Network includes an encoder and a decoder containing multiple sub-decoders; the sub-decoders contain multiple spatial feature difference processing blocks (SFDHBlock) and high-frequency feature selection blocks (HFSBlock), which, through feature difference operations and high-frequency feature extraction, specifically eliminate dust and motion blur while preserving the details of road marking edges. The specific steps are as follows:

[0065] Step S2.1: Build the encoder

[0066] Using a pre-trained UFPNet as the encoder, feature extraction is performed on the road image after grayscale value layering enhancement in step (1), outputting encoder features at 5 different scales. ,in The number of channels is The resolution is the same as the original image. .

[0067] Step S2.2: Build the decoder

[0068] Preferably, in some specific embodiments of the present invention, a decoder structure comprising two sub-decoders is constructed, each sub-decoder comprising eight (N=8) spatial feature differential processing blocks (SFDHBlock) and one high-frequency feature selection block (HFSBlock).

[0069] Step S2.2.1: Construct SFDHBlock

[0070] The Spatial Feature Difference Processing Block (SFDHBlock) includes a Spatial Feature Enhancement Module (SFEM) branch and a Simple Channel Attention (SCA) branch. The SFEM branch amplifies the feature differences between blurred and sharp regions through feature difference operations, while the SCA branch strengthens effective features through channel attention mechanisms. The two branches are fused using learnable weights to highlight the features of blurred regions and suppress background interference. The spatial feature difference operation is implemented by calculating the difference between two sets of attention weight maps and performing a weighted summation operation with the feature maps. Specific steps include:

[0071] Step S2.2.1.1: Construct the SFEM branch

[0072] SFEM highlights information in blurred regions through feature difference operations. The formula is as follows:

[0073]

[0074]

[0075] The enhanced feature map output by the SFEM branch has the same dimensions as the input features. It is used to highlight the feature information of blurred areas in the road image (such as blurred marking edges caused by dust or motion blur), providing a clear feature foundation for subsequent deblurring processing. : 1×1 convolution operation, used to adjust the number of channels in the feature map, maintaining the spatial dimension while integrating feature information between channels and avoiding dimensionality inflation; the number of convolution kernels is the same as the target number of channels, the stride is 1, and there is no padding. Reshape: Dimension reshaping function, used to reshape ( )and The processed feature matrix is ​​adjusted to meet the dimensionality requirements of a 1×1 convolution input (e.g., from "number of channels"). "Feature number" is reshaped into "channel number" high "Width", ensuring subsequent operations can be executed correctly. att1: The first set of attention weight maps, composed of the query vector. With key vector The calculated information is used to capture the correlation information of local regions in the feature map and highlight the effective features selected in the initial screening. : A learnable scalar used to adjust the weight of att2 in feature difference operations, with an initial value of 0.5; during training, it is dynamically optimized through backpropagation to control the degree of difference between att1 and att2, thereby optimizing the prominence effect of blurred region features. :att2 and The product of these values ​​is used to adjust the contribution of the second set of attention weight maps in the difference operation, avoiding the influence of bias in a single attention map on feature enhancement. att1- Feature difference operation amplifies the feature differences between blurred and clear regions by using the difference between two sets of attention weight maps, suppresses interference information in non-blurred regions (such as normal road surfaces), and focuses on the key features of blurred regions. Value vector, storing the original feature information of the feature map, along with the attention weight map ( After multiplication, attention weights can be assigned to the corresponding features to enhance the features of the fuzzy region. The attention weights are multiplied by the value vector, mapping the differencing attention weights back to the original features. This filters and enhances effective features in blurred regions while suppressing irrelevant features. SoftMax: The softmax activation function is used to adjust the attention scores. Normalize to the [0, 1] interval to ensure the sum of attention weights is 1, which facilitates subsequent feature weighting; the calculation formula is SoftMax. ,in For the first Attention score. The first query vector has the dimension "number of channels". "Feature number" is used in conjunction with the key vector. Matching calculates the degree of correlation between features and captures relevant features in ambiguous regions. Key vector The transpose operation will The dimension from "number of channels" Convert "feature count" to "feature count × channel count" to ensure consistency with... It can perform matrix multiplication operations. : Query the matrix multiplication of the key vector and its transpose to obtain the "channel number". The attention score matrix of "channel number" has each element representing the correlation strength between the corresponding channel features, reflecting the feature's ability to represent fuzzy regions. : The scaling factor for the attention score, with a value of ( (where is the number of feature channels) is used to alleviate the problem of excessively large attention scores caused by matrix multiplication, avoid the softmax function output from becoming extreme (such as one weight being close to 1 and the rest being close to 0), and improve the diversity of attention weights. : The scaled attention score matrix, through Adjusting the score range allows the softmax function to allocate attention weights more reasonably, enhancing its ability to distinguish features in fuzzy regions. The second set of query vectors and key vectors, their dimensions, functions, and... Consistent, it is used to generate a second set of attention weight maps, att2, which complements att1 and reduces the bias of single attention calculations. att2: The second set of attention weight maps works in conjunction with att1, and optimizes the salience effect of blurred region features through subsequent difference operations, thereby improving the robustness of feature enhancement. The number of feature channels, i.e. the channel dimension of the feature map (e.g., 64, 128, etc.), is determined by the output of the preceding feature extraction module (e.g., encoder) and reflects the dimensional complexity of the features.

[0076] Step S2.2.1.2: Construct the SCA branch

[0077] The SCA branch and the SFEM branch are separated by learnable weights. (Initial value 1) Fusion, the formula is:

[0078]

[0079] The fused feature map output by SFDHBlock, with dimensions similar to the input features. It is consistent and integrates the fuzzy region features of the SFEM branch with the channel attention features of the SCA branch, providing a more comprehensive feature foundation for subsequent high-frequency feature selection. Learnable weights are used to adjust the contribution of the SFEM and SCA branches in the fusion process, with an initial value of 1. During training, they are dynamically optimized through backpropagation. When the road image is severely blurred... Approaching 1 to enhance the blurred region features of the SFEM branch, when the differences in image channel features are significant, Approaching 0 to enhance the channel attention feature of the SCA branch. SFEM branch output features and weights The product of these values ​​represents the contribution of SFEM branch features to the fusion result, highlighting key features in the blurred regions. SFEM SFEM branch for input features The processing result, namely the enhanced feature of the fuzzy region obtained through feature difference operation, has a dimension of... Consistent. The weight of the SCA branch, and Complementary This ensures that the sum of the weights of the two branches is 1, avoids numerical shifts in the fused features, and maintains the stability of the feature distribution. SCA branch output features and weights The product of these two values ​​represents the contribution of the SCA branch features to the fusion result, enhancing effective features in the channel dimension (such as the channel feature differences between road markings and the road surface). SCA ( SCA branch for input features The processing result, namely the channel-weighted features obtained through a simple channel attention mechanism, has the same dimensions as... Consistent; the SCA branch enhances important channel features (such as those related to the caliper) and suppresses redundant channel features (such as background noise channels) by calculating the importance weight of each channel. Input features after layer normalization (LN) with the dimension of "number of channels". high "Width"; Layer normalization reduces the impact of feature value fluctuations on branch processing by standardizing the features of each sample to a mean of 0 and a variance of 1, thereby improving training stability and feature processing accuracy.

[0080] Step S2.2.2: Construct HFSBlock

[0081] The High-Frequency Feature Selection Block (HFSBlock) extracts high-frequency features from the features output by the last Spatial Feature Differentiation Block (SFDHBlock) using a learnable filter. Based on these high-frequency features, multiple masks with different selection ratios are generated to filter out the high-frequency feature components with the highest importance ranking. The high-frequency feature components selected by the different masks are then weighted and fused to output an enhanced high-frequency feature matrix, thereby separating and strengthening key high-frequency details related to the datum edge from the input features. Specifically, the formula for HFSBlock to extract high-frequency features using a learnable filter is as follows:

[0082]

[0083] The high-frequency feature matrix output by the feature decoupling module, with dimensions equal to the input features. Consistency is used to separate and extract high-frequency information (such as the edges of lines and the detailed features of blurred areas) in an image, providing a foundation for subsequent high-frequency feature selection. Decoupler: The feature decoupling module is a component that includes... The network components of convolution and activation functions are used to process input features. It decomposes into high-frequency features and low-frequency features, and focuses on high-frequency information that is key to deblurring by suppressing low-frequency noise (such as uniform road surface texture) and retaining high-frequency signals (such as abrupt changes in the edge of road markings). The 8th output feature of SFDHBlock, which is the fused feature after 8 spatial feature difference processing, contains rich fuzzy region features and channel attention features. It is the input basis for feature decoupling, and its dimension is "number of channels". high width".

[0084] Through the mask matrix The formula for selecting important high-frequency features is as follows:

[0085]

[0086] : The v-th mask matrix pairs input elements The processing results are used to filter the most important high-frequency features, suppress minor features, and improve feature focus. v: Mask matrix index, with values... The mask corresponds to four different filtering ratios, and the hierarchical selection of high-frequency features is achieved through multi-ratio filtering. The elements input to the mask matrix are typically the attention score matrix. The elements in the equation represent the correlation strength between high-frequency features. : Filtering rules refer to selecting the top-ranked numerical values ​​among the input elements. Elements of proportion (such as) Before selecting ; Before selecting (and so on), retaining the more critical correlation information in the high-frequency features.

[0087] When x belongs to " "When within the range, : Retain the value of this element, that is, retain the association information of the corresponding high-frequency feature.

[0088] When x does not belong to " "When within the range, Setting this element to 0 suppresses secondary correlation information of the corresponding high-frequency features, reducing redundancy.

[0089] The formula for fusing multiple mask results is:

[0090]

[0091] The high-frequency feature matrix after fusion of multiple masking results has the same dimension as the input features. It integrates effective high-frequency features under four different screening ratios, enhancing the ability to represent details in blurred areas (such as blurred parts of the caliper edge). The summation operator calculates the weighted sum of the results from processing four mask matrices, enabling the fusion of high-frequency features across multiple scales. : The learnable fusion weights of the v-th mask result, used to adjust the contribution of masks with different screening ratios in the fusion; satisfying This ensures the stability of the numerical range of the fused features; optimization is achieved through backpropagation during training, such as for severely blurred regions. Tend to increase the proportion of high screening (e.g.) Before screening The weight of ) is adjusted to retain more details. The result of the v-th mask and its weights The product of represents the contribution of the mask result to the fused features. The v-th mask matrix is ​​paired with the attention score matrix. The filtering results only retain the top-ranked ones. The attention score is set to 0, and the rest are set to 0 to highlight the important correlations between high-frequency features. : A query matrix for high-frequency features, with the dimension being "number of channels". "Eigenvalues", used in conjunction with the key matrix Matching calculates the correlation strength between high-frequency features. : Key matrix of high-frequency features, dimensions and Consistent, with Collaboratively calculate attention scores. The product of the query matrix and the transpose of the key matrix yields the "number of channels". The attention score matrix of "channel number" represents the degree of correlation between corresponding high-frequency feature channels. The higher the value, the stronger the correlation (such as the high-frequency feature channels at the edge of the caliper). : Value matrix of high-frequency features, dimensions and Consistent with the original information of high-frequency features; and consistent with the filtered attention score matrix ( After multiplication, attention weights are assigned to the corresponding high-frequency features to enhance key high-frequency information. The high-frequency features after the v-th mask are focused on by attention mechanism and mask filtering, while suppressing secondary information.

[0092] Step S2.3: Decoder Output

[0093] The decoder of the Adaptive Blurry Region Recognition Network (AIBNet) is trained using a progressive training strategy. The specific steps include: First, fixing the parameters of the shared encoder, training only the first sub-decoder by minimizing the L1 loss function between its output residual image and the residual of the real blurred-sharp image pair; then, freezing all parameters of the first sub-decoder while keeping the encoder parameters frozen, and then training the second sub-decoder, similarly minimizing its output L1 loss; this process is repeated until all sub-decoders are trained. The deblurred road image is then output using the residual image generation formula. :

[0094]

[0095] The image shows the road after it has been deblurred. For the first The residual image output by each sub-decoder; The enhanced input image.

[0096] The image recognition module 130 uses a dual attention-guided consistency network (DACNet) with an architecture that includes a shared encoder and multiple differential decoders to analyze and recognize the processed image. Each differential decoder integrates a spatial attention module and a channel attention module, and optimizes the model through a complementary consistency training strategy to output the type, wear degree and coverage area coordinate information of the road markings.

[0097] The image recognition module 130 is used to realize road image recognition. To this end, this invention proposes a Dual Attention Guided Consistency Network (DACNet) for accurately identifying the type, wear level, and location of road markings. The specific steps are as follows:

[0098] Step S3.1: Construct the DACNet model structure

[0099] Figure 3 This is a schematic diagram of the structure of the Dual Attention Guided Consistency Network (DACNet) in an embodiment of the present invention, as shown below. Figure 3 As shown, a DACNet structure is constructed, comprising one shared encoder and four differentiated decoders. The shared encoder adopts the V-Net architecture to extract features from the deblurred road image and output multi-scale semantic features. The four decoders achieve differentiation by adjusting the skip connection structure and attention module. The k-th decoder... Remove skip connections at the k-th layer and insert spatial and channel attention modules after that layer. In other words, the decoder achieves structural differentiation by removing skip connections at different levels and inserting both spatial and channel attention modules.

[0100] The spatial attention module generates spatial weights by fusing the channel average pooling and max pooling results of the feature maps, thus focusing on the spatial region where the marker line is located. Specifically, the formula for the spatial attention module is:

[0101]

[0102] Spatial attention weight map, with dimension " high Width (single-channel feature map), the value range of each element is: The weight represents the importance of the corresponding spatial location (pixel). The higher the weight, the greater the contribution of that location to road marking recognition (e.g., the weight of the area where the marking is located is higher than that of the background road surface). The sigmoid activation function normalizes the convolution output to the [0,1] interval, ensuring that the weight map can be directly used for feature weighting. Conv: The convolution operation, with a stride of 1 and padding of 1, is used to fuse the output features of average pooling and max pooling to generate a spatial attention weight map; the number of convolution kernels is 1 to ensure that the output is a single-channel weight map, integrating attention information in the spatial dimension. The feature concatenation operation concatenates the average pooling result and the max pooling result along the channel dimension to form a " high The feature map is "width"; by concatenating the two pooling results, the global statistical information (average pooling) and local saliency information (max pooling) of the feature map are captured simultaneously, improving the accuracy of spatial attention. Average pooling operation along the channel dimension is applied to the input feature map. Each spatial location Calculate the average value of all channels at this position and output " high The feature map with "width" is used to capture the global spatial distribution trend of the feature map and highlight important areas. Max pooling operation on the channel dimension is applied to the input feature map. Each spatial location Calculate the maximum value of all channels at this position and output " high The feature map of "width" is used to capture local salient signals in the feature map and highlight key local areas (such as the edges of datum lines). : The feature map input to the spatial attention module, with the dimension being "number of channels". high "Width" is usually an intermediate feature of a certain layer of the decoder, containing multi-scale semantic information of the road image (such as the outline of the road markings, the road surface background, etc.). Feature map after spatial attention enhancement, dimensions compared to the input feature map. Consistent (number of channels) high (Width) By enhancing the spatial characteristics of the road markings and suppressing interference from the background area, the feature map can improve the focus of the road markings.

[0103] Furthermore, the channel attention module generates channel weights by fusing the spatial average pooling and max pooling results of the feature maps to enhance the feature channels related to the datum lines. Specifically, the formula for the channel attention module is:

[0104]

[0105] Channel attention weight map, with the dimension being "number of channels". Each element can take values ​​ranging from 1 to 2. The weight represents the importance of the corresponding channel. The higher the weight, the greater the contribution of the channel to road marking recognition (such as channels that specifically characterize the color and edge of the markings). The sigmoid activation function, similar to the spatial attention module, is used to normalize the output of an MLP to... The interval ensures that the weight map can be directly used for channel feature weighting. MLP: Multilayer Perceptron, containing two linear layers (connected by a ReLU activation function) to perform a non-linear transformation on the pooling results, generating channel attention weights; the first linear layer compresses the feature dimension to a fraction of the original number of channels. The second linear layer restores the original number of channels, and captures the nonlinear correlation between channels through dimensionality reduction and dimensionality increase operations. The feature concatenation operation combines the average pooling result and the max pooling result in the spatial dimension along the channel dimension to form a " The feature vector of "number of channels" is obtained by concatenating the two pooling results to simultaneously capture the global statistical information (average pooling) and local extremum information (max pooling) of channel features, thereby improving the accuracy of channel attention. Average pooling operation in the spatial dimension is applied to the input feature map. For each channel, calculate the average value across all spatial locations and output the "channel count". The feature map is then flattened into a vector with the "number of channels" dimension; this is used to capture the global average response of channel features, reflecting the overall importance of the channels. Max pooling operation in the spatial dimension is applied to the input feature map. For each channel, calculate the maximum value across all spatial locations and output the "channel number". The feature map is then flattened into a vector with the dimension of "number of channels"; this is used to capture the local maximum response of channel features, reflecting the sensitivity of channels to salient features (such as datum edges). : The feature map input to the channel attention module, with the dimension being "number of channels". high "Width" is consistent with the input features of the spatial attention module and is an intermediate feature of a certain layer of the decoder. Feature map after channel attention enhancement, dimensions and input feature map Consistent (number of channels) high (Width), by strengthening channel features related to road markings and suppressing interference from redundant channels (such as noise channels), the feature map's ability to represent road markings is improved.

[0106] Step S3.2: DACNet model training

[0107] The model is optimized using complementary consistency training, and the probability map output by each decoder is... The formula for generating pseudo-tags through sharpening is as follows:

[0108]

[0109] The sharpening pseudo-labels generated by the k-th decoder for the m-th image have values ​​ranging from... This is used to enhance the confidence difference of the probability map, highlight high-confidence prediction results (such as clearly defined marking areas), and suppress interference from low-confidence, blurry areas. `m`: Image index, ranging from 1 to M (M is the total number of images), representing the m-th road image in the training set. `k`: Decoder index, with values ​​of 1, 2, 3, and 4, representing the k-th differential decoder that generates the pseudo-label. The k-th decoder for the m-th image The predicted probability map, with values ​​ranging from... This represents the probability that each pixel in the image belongs to the "road marking" category. or the probability of the "non-marking" category For the network parameters of the k-th decoder (For the parameter set). The sharpening exponent for predicting probability is calculated by... The power amplifies high-probability values ​​and compresses low-probability values, enhancing the discriminative power of the probability map; the smaller T is, the stronger the sharpening effect, with high-probability values ​​closer to 1 and low-probability values ​​closer to 0. The probability that a pixel belongs to the "non-marking" category, and Complementary (the sum of the two is 1). The sharpening exponent calculation for the probability of "non-standard lines" has the following effect: Consistency enhances the probability differentiation of "non-marked" areas. Sharpening temperature parameter, fixed value 0.1, used to control the sharpening intensity; The smaller the denominator, the more significant the stretching effect of the exponent operation on the probability, and the more obvious the difference in confidence levels of pseudo-labels, which helps to strengthen the model's learning of high-confidence regions. Normalize the probabilities of sharpened "markers" and "non-markers" to ensure that false labels are not detected. The range of values ​​is It conforms to the characteristics of a probability distribution.

[0110] Furthermore, when training the Dual Attention Guided Consistency Network (DACNet), a complementary consistency training strategy is employed. This strategy includes: sharpening the prediction results output by each differential decoder to generate high-confidence pseudo-labels; and optimizing network parameters by calculating the consistency loss between the prediction results of different differential decoders, thereby improving the stability and generalization ability of the model's predictions. Specifically, the consistency loss between different differential decoders is calculated using the mean squared error (MSE), as shown in the formula:

[0111]

[0112] Unsupervised complementary consistency loss is used to measure the consistency of the prediction results of the four differential decoders. By minimizing this loss, the line recognition results of different differential decoders on the same image tend to be consistent, thereby improving the robustness and generalization ability of the model. The loss is summed across all M images in the training set to reflect the consistency across the entire dataset. M: The number of images in the training set, i.e., the total number of road images used in the training, which determines the sample size for loss calculation. The loss of all different decoder pairs (j and k) in the four decoders is accumulated (4×3=12 pairs in total) to ensure that the complementary relationship between all decoders is covered and to avoid the influence of the bias of a single decoder pair. : Decoder index, j represents the current decoder for prediction, k represents the decoder that generated the pseudo-labels, Ensure consistency across different decoders (rather than their own consistency). MSE: Mean Squared Error Function, used to quantify the difference between two probability distributions. Here, it is used to calculate the difference between the predicted probability of decoder j and the sharpened pseudo-label of decoder k. The smaller the value, the higher the consistency. The j-th decoder for the m-th fused input image The predicted probability map, with dimensions consistent with the image pixels, reflects the decoder's... Determine the marked areas in the image. The fused version of the m-th input image, typically a deblurred road image, serves as the unified input for all decoders, ensuring that predictions from different decoders are based on the same image data and fairly measuring consistency.

[0113] Step S3.3: Model Output

[0114] The probability map output by the decoder is thresholded (threshold 0.7) and morphologically processed to obtain a binary map of the marking region. Combined with a pre-trained marking type classifier (based on ResNet-50) and a wear level assessment model, the marking type, wear level and coverage area coordinates are output.

[0115] The wear assessment model is based on the difference in average grayscale values ​​between the marking area and the adjacent standard road surface area. Using the smallest bounding rectangle of the marking area polygon as a base, a ring-shaped region is formed by extending it outwards by a specific width (e.g., 20 pixels). This ring-shaped region is defined as the background (adjacent standard road surface) area for comparison. The difference between the average grayscale value of pixels within the marking area and the average grayscale value of pixels in the surrounding normal road surface area is calculated. Based on the preset range of this difference, the wear level is classified as mild, moderate, or severe. For example, a difference greater than 50 indicates mild wear, between 20 and 50 indicates moderate wear, and less than 20 indicates severe wear.

[0116] The parameter decision module 140 is used to adaptively match and adjust the wire removal pressure and operating speed parameters of the pneumatic wire removal machine based on the results output by the image recognition module.

[0117] The parameter decision module 140 is used to implement parameter decisions. It has a built-in mapping table between the marking condition and the removal parameters, and uses a fuzzy PID control algorithm to dynamically calculate the adjustment amount for the removal pressure or operating speed based on the deviation between the currently identified marking condition and the preset standard value. Specifically:

[0118] In this embodiment of the invention, a mapping table of marking conditions and removal parameters is established. For example, a lightly worn solid line marking corresponds to a removal pressure of 1.0 MPa and an operating speed. The dashed line indicating moderate wear corresponds to a pressure of 1.2 MPa and a speed of Combining real-time road marking information, the parameters are dynamically adjusted using a fuzzy PID algorithm, with the following formula:

[0119]

[0120] The control output at time t represents the real-time adjustment of the wire removal machine's wire removal pressure or operating speed; if it's a pressure adjustment, the unit is MPa; if it's a speed adjustment, the unit is... The positive or negative sign of the adjustment amount represents whether the parameter is "increased" or "decreased," used to dynamically correct the deviation between the current parameter and the ideal value. t: a time variable, representing the moment of the current parameter adjustment (such as a second-level time node in the operation process), ensuring the real-time nature of the parameter adjustment and adapting to the dynamic changes in road marking conditions. : Scale factor, fixed value of 0.8, used to amplify the current deviation. Impact on control output; The larger the value, the more sensitive the deviation response, but this may lead to parameter oscillations; =0.8 is the optimal value that balances response speed and stability, and is suitable for road marking removal scenarios. The deviation value at time t is the difference between the average gray value difference between the current marking area and the road surface, which is evaluated in real time by the image recognition module 130, and the preset standard gray value difference threshold for the corresponding marking type. For example, for a lightly worn solid line, the standard gray value difference threshold is 30. If the gray value difference calculated by the current recognition is 25, then e(t) = 30 - 25 = 5 (the deviation is positive, indicating that the current line removal effect is insufficient and the parameters need to be adjusted to enhance the line removal effect). Integral coefficient The integration time is a fixed value of 1.2. Used to adjust the contribution of the integral term to the control output; the integral term is calculated by accumulating historical deviations. This eliminates long-standing static deviations (such as insufficient parameters in areas of continuous mild wear). The larger the value, the weaker the integral effect, thus avoiding parameter overshoot caused by integral saturation. : The integral of the deviation from the initial time (0) to the current time (t), that is, the cumulative sum of historical deviations; used to capture long-term deviation trends, for example, if the wear of the marking is detected to be higher than the standard value for 10 consecutive seconds, the integral term will accumulate the deviation, driving the control output to continuously adjust until the deviation is eliminated. The derivative time, with a fixed value of 0.3, is used to enhance the response of the control output to the rate of change of the deviation. The derivative term suppresses the oscillation of parameter adjustment and improves system stability by predicting the trend of deviation change (such as increasing the adjustment amount in advance when the deviation increases rapidly). The rate of change of the deviation at time t is the derivative of the deviation with time; if This indicates that the deviation is "increasing" (e.g., a sudden increase in wear), and the derivative term will drive the control output to increase the adjustment amount in advance; if If the deviation is "decreasing", then the adjustment amount should be reduced appropriately to avoid over-correction of the parameters.

[0121] The formula as a whole achieves smooth and precise adjustment of the line marking parameters by combining a "proportional term (responding to the current deviation) + integral term (eliminating historical deviations) + derivative term (predicting future deviations)". This avoids the static deviation of single proportional control and prevents parameter oscillations, adapting to the complex changes in road marking conditions.

[0122] The execution control module 150 is used to control the actuator of the wire removal machine according to the output signal of the parameter decision module, including a pressure regulating valve and a walking drive motor, so as to realize closed-loop control of the wire removal parameters.

[0123] The execution control module 150 is used to implement execution control. Specifically, in this embodiment of the invention, after receiving parameter signals, the execution control module 150 adjusts the high-pressure airflow pressure (range 0.8-1.5MPa) via an electromagnetic proportional valve, with an accuracy of ±0.05MPa; the walking drive motor adjusts the working speed (range 3-8m / min) via frequency conversion speed regulation, with an accuracy of ±0.2m / min, and simultaneously feeds back the actual parameters to the parameter decision module in real time to achieve closed-loop control.

[0124] The intelligent operation system 100 for pneumatic road marking removal based on AI image recognition provided in the above embodiments of the present invention employs grayscale layering and illumination-adaptive Gamma correction technology to significantly improve the contrast of road marking features under different lighting conditions; based on feature difference and high-frequency selection adaptive deblurring network (AIBNet), it accurately eliminates blurring while preserving the edge details of the road markings; and it constructs a recognition network (DACNet) integrating a dual attention mechanism and a consistency training strategy, which greatly improves the recognition accuracy and robustness of road marking type, wear degree, and location. Ultimately, the present invention can dynamically adjust the removal parameters according to the real-time identified road marking conditions through a fuzzy PID control algorithm, achieving efficient, accurate, and adaptive road marking removal operations, effectively avoiding road surface damage, and significantly reducing the cost of manual intervention.

[0125] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0126] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0127] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. An intelligent operating system for a pneumatic road marking removal machine based on AI image recognition, characterized in that, include: The image acquisition module is used to acquire original images containing road markings and the ambient light intensity value for each image. The image processing module is used to sequentially perform contrast enhancement and deblurring processing on the original image to improve the clarity and distinguishability of the marking features, thereby obtaining the processed image. The image recognition module uses a dual attention-guided consistency network (DACNet) with an architecture that includes a shared encoder and multiple differential decoders to analyze and recognize the processed images. Each differential decoder integrates a spatial attention module and a channel attention module, and optimizes the model through a complementary consistency training strategy to output the type, wear level and coverage area coordinates of road markings. The parameter decision module is used to adaptively match and adjust the wire removal pressure and operating speed parameters of the pneumatic wire removal machine based on the results output by the image recognition module. The execution control module is used to control the actuator of the wire removal machine according to the output signal of the parameter decision module, including a pressure regulating valve and a walking drive motor, to realize closed-loop control of the wire removal parameters.

2. The system according to claim 1, characterized in that, The image processing module includes an image enhancement unit and an image deblurring unit; wherein... The image enhancement unit is used to enhance the contrast of the original image. The enhancement process includes: decomposing the image using grayscale layering technology, and performing enhancement operations combining histogram truncation and Gamma correction on the decomposed illumination components to obtain the enhanced road image; wherein, the parameters of the Gamma correction are dynamically configured according to the ambient light intensity value to adaptively adjust the enhancement processing strategy. The image deblurring unit is implemented using an Adaptive Blur Region Recognition Network (AIBNet) to deblur the enhanced road image. The Adaptive Blur Region Recognition Network includes an encoder and a decoder containing multiple sub-decoders. The sub-decoders contain multiple Spatial Feature Differentiation Blocks (SFDHBlock) and High Frequency Feature Selection Blocks (HFSBlock). Through feature difference operations and high frequency feature extraction, dust and motion blur are specifically eliminated while preserving the details of the road marking edges.

3. The system according to claim 2, characterized in that, In the image enhancement unit, the grayscale value layering technique specifically includes: Perform a logarithmic transformation on the original image to decompose it into an illuminated image and a reflected image; A vector grid corresponding to the original image is constructed, and the vector grid is smoothed by a three-dimensional Gaussian filter; Bilateral filtering is used to further optimize the filtered vector grid, preserving image edge details.

4. The system according to claim 2, characterized in that, The parameters for Gamma correction include a brightness gain parameter 'a' and a contrast control parameter 'd'; the dynamic configuration involves assigning different combinations of values ​​to the brightness gain parameter 'a' and the contrast control parameter 'd' based on a preset range of ambient light intensity values, specifically including: When the ambient light intensity value is less than or equal to the first preset threshold, the image enhancement unit is configured to use a parameter combination of a>1 and d<1 to improve the overall brightness of the image and enhance the contrast of dark areas. When the ambient light intensity value is greater than or equal to the second preset threshold, the image enhancement unit is configured to use a parameter combination of a<1 and d>1 to reduce the overall brightness of the image and restore the details in the bright areas; Wherein, the first preset threshold is less than the second preset threshold.

5. The system according to claim 2, characterized in that, In the image deblurring unit, the spatial feature difference processing block (SFDHBlock) includes a spatial feature enhancement module (SFEM) branch and a simple channel attention (SCA) branch; The Spatial Feature Enhancement (SFEM) branch amplifies the feature differences between blurred and clear regions through feature difference operations, while the Simple Channel Attention (SCA) branch strengthens effective features through channel attention mechanisms. The two branches are fused through learnable weights to highlight the features of blurred regions and suppress background interference. The spatial feature difference budget is achieved by calculating the difference between two sets of attention weight maps and performing weighted operations with the feature maps.

6. The system according to claim 2 or 5, characterized in that, In the image deblurring unit, the high-frequency feature selection block (HFSBlock) extracts high-frequency features from the features output by the last spatial feature difference processing block (SFDHBlock) using a learnable filter. Based on the high-frequency features, multiple masks with different selection ratios are generated to select high-frequency feature components with higher importance. The high-frequency feature components selected by the different masks are weighted and fused to output an enhanced high-frequency feature matrix, thereby separating and enhancing key high-frequency details related to the datum edge from the input features.

7. The system according to claim 1, characterized in that, The differential decoder achieves structural differentiation by removing skip connections at different levels and inserting spatial and channel dual attention modules; The spatial attention module generates spatial weights by fusing the channel average pooling and max pooling results of the feature map, so as to focus on the spatial region where the datum line is located. The channel attention module generates channel weights by fusing the spatial average pooling and max pooling results of the feature maps, thereby strengthening the feature channels related to the datum lines.

8. The system according to claim 7, characterized in that, When training the dual-attention guided consistency network (DACNet), a complementary consistency training strategy is employed, which includes: The prediction results output by each of the differential decoders are sharpened to generate high-confidence pseudo-labels; By calculating the consistency loss between prediction results from different differential decoders, network parameters can be optimized to improve the stability and generalization ability of model predictions.

9. The system according to claim 1, characterized in that, The control decision module has a built-in mapping table between the marking status and the line removal parameters, and adopts a fuzzy PID control algorithm to dynamically calculate the adjustment amount of the line removal pressure or operation speed based on the deviation between the currently identified marking status and the preset standard value.

10. The system according to claim 1, characterized in that, The system also includes a lidar module, used to synchronously acquire three-dimensional point cloud data and reflectivity information of the road surface with the image acquisition module, as auxiliary data, for cross-validation or supplementary judgment of the road marking recognition results in complex scenarios.

Images