Wear-resistant pipe surface wear detection method and system based on image processing

Abstract

The application belongs to the technical field of image processing, and particularly relates to a wear-resistant pipe surface wear detection method and system based on image processing, which comprises the following steps: acquiring a to-be-detected image of a wear-resistant pipe surface, inputting the to-be-detected image into a multi-layer convolution feature extraction network for layered feature extraction to obtain feature maps of each layer; and performing nonlinear activation processing on pre-activation values of each neuron in the feature maps output by at least one convolution layer of the multi-layer convolution feature extraction network, specifically: when the pre-activation value is greater than zero, the output of the neuron is the pre-activation value; and when the pre-activation value is less than or equal to zero, the output of the neuron is the product of the pre-activation value and a complex negative slope coefficient. The application can enhance the recognition and extraction capability of the model for small and irregular wear features under a complex texture background of the wear-resistant pipe surface, and improve the accuracy of the entire wear detection method.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a method and system for detecting wear on the surface of wear-resistant pipes based on image processing. Background Technology

[0002] During industrial production, the inner walls of pipelines are inevitably subjected to high-speed scouring, intense friction, and chemical corrosion from various materials, resulting in various physical wear phenomena, including surface scratches, metal pits, and coating peeling. This wear not only significantly reduces fluid transport efficiency but, in extreme cases, can further deteriorate into severe pipe wall perforation or structural fracture, leading to serious safety accidents and incalculable economic losses. Therefore, regular, high-precision inspection and condition assessment of the physical wear condition of wear-resistant pipe surfaces are of crucial engineering significance for ensuring industrial production safety, scientifically predicting the service life of production equipment, and rationally formulating equipment downtime maintenance plans.

[0003] In current industrial practice, traditional surface inspection methods mainly rely on manual visual inspection or basic image processing techniques. This is not only inefficient and highly dependent on the subjective experience of the inspectors, but also poses significant personal safety risks in harsh industrial environments. When using basic image processing techniques, algorithm engineers typically need to manually plan and design feature extraction operators and edge detection algorithms. These manually designed feature extraction algorithms are extremely sensitive to environmental interference factors such as complex lighting conditions in industrial environments, specular reflections on metal surfaces, and complex background textures and noise. This results in insufficient overall system stability and inadequate generalization ability to adapt to different types of wear.

[0004] To address the limitations of traditional visual algorithms, deep learning technology has been gradually introduced into the field of surface defect detection. Multilayer convolutional feature extraction networks can directly learn and extract deep, high-dimensional semantic information about pipe surface wear features from the original two-dimensional pixel image matrix in an end-to-end manner, greatly improving the accuracy of automated detection tasks. In a typical multilayer convolutional architecture design, the nonlinear activation function module plays a core role in injecting nonlinear mapping transformation capabilities into the entire linear convolutional network system. It is one of the key computational links that ultimately determines the overall feature extraction performance and detection accuracy of the model. Currently, the widely used modified linear unit activation function sets all negative pre-activation values ​​directly to zero, which may lead to the permanent loss of some effective information, including weak wear signals, during feature propagation, which is not conducive to detecting small defects in the early stages of wear.

[0005] However, while the LeakyReLU activation function alleviates the information loss problem by introducing a fixed negative slope, this slope is globally uniform and cannot be differentiated according to the semantic importance of different spatial locations in the feature map and the structural complexity of local regions. For targets like wear, which have random spatial distribution, rich texture details, and diverse defect morphologies, existing activation functions cannot guide the network to enhance the response of key wear regions while suppressing irrelevant background noise, thus affecting the model's feature representation ability and final detection accuracy. Summary of the Invention

[0006] To address the technical problems in existing technologies where the loss of feature information and the inability to differentiate response spatial location and local structural information due to activation function processing mechanisms affect the accuracy of wear detection, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a method for detecting wear on the surface of a wear-resistant pipe based on image processing, comprising: acquiring an image to be detected of the surface of the wear-resistant pipe, and inputting the image to be detected into a multi-layer convolutional feature extraction network for hierarchical feature extraction to obtain feature maps of each layer; after at least one convolutional layer of the multi-layer convolutional feature extraction network, performing nonlinear activation processing on the pre-activation value of each neuron in the feature map output by the convolutional layer, specifically: when the pre-activation value is greater than zero, the output of the neuron is the pre-activation value; when the pre-activation value is less than or equal to zero, the output of the neuron is the product of the pre-activation value and the composite negative slope coefficient. The composite negative slope coefficient is calculated based on spatial adjustment parameters, local structure factors, and basic leakage coefficient, and then truncated to a range of zero to one. The spatial adjustment parameters are parameter values ​​in the trainable spatial sensitivity matrix corresponding to the neuron's position. The local structure factor is a local feature dispersion index calculated based on the local neighborhood features centered on the neuron in the current layer feature map. The basic leakage coefficient is a preset positive hyperparameter. Based on the feature representation obtained after nonlinear activation processing, the wear characteristics of the wear-resistant pipe surface are determined, including wear regions or wear levels.

[0008] This invention dynamically integrates the absolute physical space sensitivity of neuron nodes with the microscopic texture structure complexity of local images, and superimposes an indispensable underlying additive bias. This enables each neuron node in the network hardware loop to have an adaptive and dynamically changing activation response characteristic based on its own physical coordinates and the content being processed. Thus, while effectively suppressing and smoothing background noise, it can perfectly preserve and differentially amplify those weak negative physical features that are crucial for early wear identification. From the underlying mathematical deduction, it eliminates the gradient vanishing problem of weak electrical signals and neuron inactivation, significantly improving the accuracy of the measurement system in identifying and extracting extremely small and irregular primary wear features against complex industrial texture backgrounds.

[0009] Preferably, the configuration method of the trainable spatial sensitivity matrix includes: initializing all elements in the spatial sensitivity matrix to preset positive constants before model training begins; and performing gradient calculation and parameter update on the spatial sensitivity matrix through backpropagation algorithm combined with a set loss function during model training.

[0010] This invention ensures that the hardware network has a physically consistent basic level adjustment capability in the initial startup state by globally initializing the spatial sensitivity matrix to a positive constant. Furthermore, the dynamic physical gradient update through the backpropagation algorithm endows the two-dimensional matrix with a powerful spatial physical evolution capability, enabling it to remember the high-frequency and frequent areas of physical wear on the inner wall of the pipe. In subsequent real-time industrial online inference detection, it can automatically assign higher weight gain to weak abnormal electrical signals at key physical spatial locations, greatly improving the targeting accuracy and recall rate of defect detection.

[0011] Preferably, the acquisition of the local structure factor includes: calculating the local variance or local standard deviation of the current layer feature map using a neighborhood window of a preset size, generating a local discreteness feature map of the same size as the feature map in real time, obtaining the feature value corresponding to the neuron coordinates based on the real-time calculated local discreteness feature map, and normalizing the feature value to the interval between zero and one through a linear mapping function to obtain the local structure factor.

[0012] This invention uses a blind-zone-free sliding scanning window to calculate microscopic local variance in real time. The data processing system can accurately and dynamically sense the roughness of the physical region where the current electrical signal pixel is located. In the region of a smooth and intact metal pipe wall, the calculated local structure factor is greatly reduced, thereby effectively cutting off and suppressing the interference transmission of common-mode physical noise. In the physical boundary edge region where there are mechanical scratches or material peeling, the local structure factor rises sharply, guiding the subsequent calculation loop to retain the multiplied micro-boundary texture features, thus enhancing the system's high-fidelity capture capability of complex micro-damage boundary morphology.

[0013] Preferably, the basic leakage coefficient is a constant with a value between zero and one.

[0014] Preferably, the step of acquiring the image to be detected on the surface of the wear-resistant pipe includes: uniformly illuminating the surface of the wear-resistant pipe to be detected using polarized light or diffuse reflection light source; acquiring the original image of the surface of the wear-resistant pipe using image acquisition hardware; adjusting the pixel size of the original image to match the size of the preset input dimension of the multi-layer convolutional feature extraction network, and performing numerical normalization operations on the mean and standard deviation of the pixel matrix of the original image to obtain the image to be detected.

[0015] This invention employs a precision orthogonal polarization optical path system at the hardware front end, thoroughly filtering out the blinding specular high-reflection photons easily generated by the physical curved surface of the metal cylinder from the physical optical source. This ensures that the light signal ultimately entering the photoelectric sensor carries the true physical diffuse reflection information of the wear pits to the greatest extent. The subsequent zero-mean normalized tensor operation eliminates the system polarization error introduced by the drastic fluctuations in background lighting intensity at different working times and batches in the industrial field from a purely mathematical perspective. This anchors the messy image data uniformly within the physical level range of the standard normal distribution. This greatly accelerates the convergence speed of subsequent deep matrix operations and ensures the stability of the entire system signal reference.

[0016] Preferably, when the wear feature is a wear region, determining the wear feature of the wear-resistant pipe surface includes: inputting the feature representation into the decoder feature reconstruction network; by performing at least one upsampling or deconvolution calculation operation, gradually restoring the spatial feature size of the feature map to a predetermined size consistent with the image to be detected, generating a pixel-level feature prediction map; applying a nonlinear probability mapping function to each pixel feature value in the feature prediction map to calculate the probability calibration value of each independent pixel belonging to the wear region; comparing the probability calibration value of each independent pixel with a preset judgment threshold to determine whether each independent pixel belongs to the wear region, thereby generating a binary mask map composed of background pixels and wear pixels to obtain the wear region on the wear-resistant pipe surface.

[0017] This invention constructs a hardware architecture for hierarchical spatial expansion physical reverse mapping of the decoder network. Combined with a rigorous logarithmic probability output mapping function and a high-precision hard-decision threshold voltage comparator, the system can reproject and restore the high-order electrical feature matrix, which is in an extremely abstract dimension, into a two-dimensional coordinate system image with extremely high physical resolution. The final output binary digital mask matrix effectively removes the blurred transition grayscale, directly providing the display and control terminal of the industrial control system or the mechanical maintenance arm with the geometric area and coordinate vector data of the wear patch with absolutely clear physical boundaries, truly realizing precise positioning for industrial field applications.

[0018] Preferably, the computation of the decoder feature reconstruction network includes: after each upsampling or deconvolution operation, extracting shallow feature maps with the same physical size from the multi-layer convolutional feature extraction network; and merging the shallow feature maps with the new feature maps obtained after the upsampling or deconvolution operation in the channel dimension to form a fused map with multi-scale features.

[0019] Preferably, when the wear feature is a wear level, determining the wear feature of the wear-resistant pipe surface includes: performing a global spatial dimension pooling operation on the feature representation to convert the multi-channel two-dimensional feature matrix into a one-dimensional feature vector of fixed length; inputting the converted one-dimensional feature vector into a feature mapping structure containing at least one fully connected layer for feature space transformation and dimension compression adjustment; setting the number of output computing nodes of the last fully connected layer to be equal to the preset number of wear level classifications; applying a probability distribution normalization mapping function to the feature vector output by the last fully connected layer to calculate the probability distribution values ​​of the current image to be detected belonging to each preset wear level, and extracting the category label corresponding to the maximum probability distribution value as the final determined wear level.

[0020] Preferably, obtaining the composite negative slope coefficient specifically includes: multiplying the local structure factor by the spatial adjustment parameter at the corresponding position to form a dynamic adjustment term, performing an additive bias operation on the dynamic adjustment term and the preset basic leakage coefficient to obtain a preliminary coefficient, and restricting the preliminary coefficient to the interval between zero and one by a truncation function to obtain the final composite negative slope coefficient.

[0021] Secondly, the present invention provides a wear-resistant pipe surface wear detection system based on image processing, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned wear-resistant pipe surface wear detection method based on image processing is implemented.

[0022] By adopting the above technical solution, the above-mentioned image processing-based wear detection method for wear-resistant pipe surfaces is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0023] The beneficial effects of this invention are as follows: This invention provides a wear detection method and system for wear-resistant pipe surfaces based on image processing. It not only solves the core problem of weak damage electrical signals being easily truncated due to negative values ​​in traditional algorithms from a mathematical and physical perspective, but also endows the detection system with the significant advantage of maintaining extremely high detection sensitivity and stable anti-interference ability when facing extremely harsh industrial light pollution and highly complex metal reflection textures. It provides a solid, reliable and clearly defined measurement means for safety early warning and predictive maintenance in industrial sites. Attached Figure Description

[0024] Figure 1 This is a flowchart of the wear detection method for wear-resistant pipe surfaces based on image processing according to the present invention; Figure 2 This is a comparison chart of the mapping curves of the present invention and traditional activation functions. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention discloses an image processing-based method for detecting wear on the surface of wear-resistant pipes, referring to... Figure 1 This includes steps S1-S4: S1. Obtain the image to be detected on the surface of the wear-resistant pipe, and input the image to be detected into a multi-layer convolutional feature extraction network for hierarchical feature extraction to obtain the feature map of each layer.

[0028] In an optional embodiment, a linear CCD camera is used to scan and acquire grayscale images along the axial or circumferential direction of the wear-resistant pipe. After scanning and stitching, a grayscale image with a resolution of 2048 pixels × 2048 pixels is obtained as the original image. An image processing library is called to normalize the original image by subtracting the mean of all pixels in the image pixel by pixel and then dividing by the standard deviation. The tensor of the preprocessed original image is input into a pre-trained multi-layer convolutional feature extraction network such as ResNet50 or VGG16. Through the forward propagation process of the model, convolution, pooling and other operations are performed layer by layer. The output of each convolutional layer is the feature map of that layer. During image acquisition, a cross-polarized illumination scheme is employed. Specifically, a linear polarizer is installed in front of the main light source, and another linear polarizer is installed in front of the lens of the industrial camera, with the polarization directions of the two polarizers perpendicular to each other. This physical optical structure eliminates specular highlights on the metal surface of the wear-resistant tube, thereby highlighting the diffuse reflection area caused by wear. The camera aperture is set to F8 to ensure sufficient physical depth of field to cover the curved tube wall, the shutter speed is set to 1 / 2000 second to prevent motion blur caused by physical scanning motion, and the ISO is set to 100 to minimize image noise.

[0029] After acquiring the original image, a size scaling operation is performed. The central processing unit calls the size scaling function in the image processing library. The scaling function uses a bicubic interpolation algorithm to uniformly scale the size of the original image to the input size required by the multi-layer convolutional feature extraction network. In this embodiment, it is scaled to a standard input size matrix with a height of 512 pixels and a width W of 512 pixels.

[0030] The central processing unit then performs numerical normalization on the 512×512 pixel matrix, following the Z-score normalization physical relation: .

[0031] In the formula, This indicates that after normalization, the value located at the th position... line, number The value of the pixel in the column. This indicates that within a 515×512 pixel matrix, the physical two-dimensional spatial coordinates are located at the [missing information]. line, number The original grayscale value of that individual pixel in the column. This represents the arithmetic mean of the original grayscale values ​​of all 262,144 pixels in the entire 515×512 pixel matrix. This represents the global standard deviation of the original grayscale values ​​of all 262,144 pixels in the pixel matrix.

[0032] For example, the system traverses the entire pixel matrix and calculates the arithmetic mean of the current image to be detected. The global standard deviation is 128.5. The value is 45.2, and the coordinates extracted from the pixel matrix by the system are... The target pixel is read to obtain its original grayscale value. Substituting into the above formula, we get: After pixel-by-pixel traversal operations by the central processing unit, the original image pixel matrix is ​​converted into a floating-point tensor matrix with a mean of 0 and a standard deviation of 1.

[0033] The system inputs the preprocessed floating-point tensor matrix into a pre-constructed multi-layer convolutional feature extraction network. In this embodiment, the multi-layer convolutional feature extraction network specifically adopts the ResNet50 feature extraction backbone architecture. The tensor matrix sequentially passes through multiple two-dimensional physical convolutional layers within the feature extraction backbone architecture. In each convolutional layer, the system uses a pre-sized convolutional kernel weight matrix to perform a sliding space inner product operation with the input data, thereby outputting feature map tensors representing different deep physical semantics at each layer.

[0034] S2. After at least one convolutional layer in the multi-layer convolutional feature extraction network, the pre-activation value of each neuron in the feature map output by the convolutional layer is subjected to non-linear activation processing. Specifically, when the pre-activation value is greater than zero, the output of the neuron is the pre-activation value; when the pre-activation value is less than or equal to zero, the output of the neuron is the product of the pre-activation value and the composite negative slope coefficient.

[0035] In an optional embodiment, after a specific layer of the multi-layer convolutional feature extraction network, the system extracts the feature map matrix output by that layer, and pre-activation values ​​for each individual neuron node in the feature map matrix. The system executes non-linear hardware logic judgments and arithmetic branches: When the detector recognizes the pre-activation value At this time, the output channel of the neuron node directly outputs the original value of the pre-activation value.

[0036] When the detector recognizes the pre-activation value At that time, the output channel of the neuron node outputs the value of the pre-activation value and a composite negative slope coefficient calculated in real time.

[0037] S3, the composite negative slope coefficient is calculated based on the spatial adjustment parameter, local structure factor and basic leakage coefficient, and then truncated to be limited to the range of zero to one; the spatial adjustment parameter is the parameter value corresponding to the neuron position in the trainable spatial sensitivity matrix; the local structure factor is the local feature dispersion index calculated based on the local neighborhood features centered on the neuron in the current layer feature map; the basic leakage coefficient is a preset positive hyperparameter.

[0038] In an optional embodiment, for the pre-activation value To prevent gradient vanishing in extreme regions, the system must ensure that each neuron node has basic leakage conductivity. The composite negative slope coefficient is calculated using the following formula. : ; For weak electrical signal characteristics located on the negative half-axis, the system does not use the traditional direct zeroing operation, but instead calculates a dynamic attenuation ratio between 0 and 1. This ratio is determined by the spatial absolute position weight of the target pixel, the intensity of the physical texture around the target pixel, and the system's bottom-line guarantee constant.

[0039] In the formula, To represent spatial adjustment parameters, a readable and writable data area with dimensions identical in two dimensions to the currently processed feature map is allocated in the system memory and stored as a two-dimensional array named Spatial Sensitivity Matrix S. At the initial training stage of the multi-layer convolutional feature extraction network, all elements in Spatial Sensitivity Matrix S are forcibly initialized to a preset positive constant of 0.01. During long-term training, gradient calculations and parameter updates are performed on the Spatial Sensitivity Matrix using the backpropagation algorithm combined with a set loss function. The set loss function can be a binary cross-entropy loss function or a mean squared error loss function. The values ​​within Spatial Sensitivity Matrix S are continuously updated, thus remembering which coordinate positions in the image are more prone to wear. For the current feature map, the two-dimensional spatial physical coordinates at position i are... line, number For each neuron node in the column, the system directly retrieves the parameter values ​​at the corresponding coordinate positions from the spatial sensitivity matrix S. Subsequently, the system performs absolute value mathematical operations on the read values ​​to obtain the spatial adjustment parameters. .

[0040] The local structure factor characterizes the texture roughness of the micro-region where the current pixel is located. If the acquired region is a smooth, intact pipe wall, the structure factor is extremely low; if the acquired region has peeling physical edges, the structure factor is extremely high. The system uses the coordinates of the neuron node on the current layer's feature map plane. To determine the absolute geometric center, a physical neighborhood window with a preset parameter of K×K is used; in this embodiment, a 3×3 physical neighborhood window is used. The system captures the floating-point feature values ​​of these 9 pixel nodes. The system calculates the local mathematical standard deviation for these 9 floating-point feature values. Then, the system normalizes the calculated local standard deviation value to a closed interval of 0 to 1 using a linear mapping function. This normalized dimensionless real number is the local structure factor. .

[0041] Specifically, obtain the maximum local standard deviation of all local windows in the current layer feature map. The minimum local standard deviation of all local windows in the current layer feature map. Using relational expressions Perform calculations, where This represents the calculated local standard deviation within the local neighborhood window of the current target neuron, or it can be truncated and normalized by setting a fixed empirical maximum value.

[0042] The basic leakage coefficient is a constant between zero and one. In this embodiment, The value is 0.05, thus providing an absolute non-zero mathematical baseline. This ensures that even if both the spatial adjustment parameter and the local structure factor approach absolute zero, the neuron still retains a weak gradient conduction capability on the negative half-axis, effectively avoiding the gradient vanishing and neuron inactivation problems during the training process of multi-layer convolutional feature extraction networks.

[0043] This is a cutoff function, a physical hardware limiter. Its input is the mixed output value obtained by performing an additive bias operation on the product of the aforementioned dynamic adjustment term and the basic leakage coefficient. If the mixed output value is less than 0, the limiter outputs 0; if the mixed output value is greater than 1, the limiter outputs 1; if the mixed output value is greater than or equal to 0 and less than or equal to 1, the limiter outputs the mixed output value as is, ensuring that the final composite negative slope coefficient is always strictly clamped within an absolutely safe physical level multiple range.

[0044] For example, in the current feature map, the coordinates are located in the 10th row and 15th column, that is... The pre-activation values ​​of the neuron nodes output by the preceding convolution. The value is -2. The spatial sensitivity matrix is ​​read by addressing, and the result is obtained. The absolute value spatial adjustment parameter was calculated to be 0.8, and the coordinate system was used as the reference. The standard deviation of the 3×3 neighborhood values ​​is calculated and normalized to obtain the local structure factor. The value is 0.5, and the statically set baseline leakage coefficient is retrieved. The value is 0.05. Substitute it into the formula to calculate: Finally, the output value of this neuron node is calculated: The output value is stored in the physical input buffer of the next layer network. The system repeats all the above operation steps for the coordinates of each independent neuron in the feature map of the current layer.

[0045] S4. Based on the feature representation obtained after nonlinear activation processing, determine the wear characteristics of the wear-resistant pipe surface. The wear characteristics include wear area or wear level.

[0046] In an optional embodiment, the system inputs the processed feature representation tensor into a decoder network with skip connections. The decoder network contains multiple levels of physical upsampling modules. In each upsampling module, the system performs a deconvolution matrix operation, setting the deconvolution kernel size to 3×3 and the stride parameter to 2. Each deconvolution operation physically magnifies the length and width of the input feature map tensor in two-dimensional space by a factor of two. Simultaneously, through hardware skip connections, feature maps with the same physical resolution are directly retrieved from the encoder and stitched together along the channel dimension to compensate for the spatial high-frequency physical details lost in deeper networks. After multiple levels of deconvolution and fusion operations, the decoder outputs a 512×512 pixel-level prediction map matrix with the exact same size as the original input image.

[0047] For each independent pixel in the 512×512 prediction image matrix, the system applies the Sigmoid log-odds output function to each pixel's output value. In the formula, These are the original values ​​in the prediction graph. Through a physical transformation of the exponential base, all values ​​are mapped to a probability scaling value that is strictly within the open interval of 0 to 1. This probability calibration value This indicates the probability that the corresponding pixel belongs to the wear pit area of ​​the metal pipe wall.

[0048] The system internally stores an unchangeable preset judgment threshold. The system uses a hardware comparator to compare and determine each of the 262,144 probability scaling values ​​in the matrix: if the probability scaling value of a certain coordinate is greater than the determination threshold of 0.5, the pixel value of the corresponding coordinate is assigned a value of 1 in a brand new, equal-sized mask image; if the probability scaling value of a certain coordinate is less than or equal to the determination threshold of 0.5, the pixel value of the corresponding coordinate in the mask image is assigned a value of 0. The final generated binary digital mask image, containing only 0s and 1s, is the final output result that physically represents the location and geometry of the wear area on the inner wall of the pipe.

[0049] In an optional embodiment, for the deep feature representation tensor that has undergone nonlinear activation processing, for example, assuming its physical dimensions are: number of channels C=512, height H=16, width W=16, the system performs a global two-dimensional average pooling operation on the feature representation. Specifically, for each independent 16×16 matrix surface in the 512 channels, the 256 floating-point values ​​inside are arithmetically summed and divided by 256 to obtain a single average value. Through this operation, the original three-dimensional feature tensor is compressed to a one-dimensional physical feature vector with a strict length of 512 floating-point numbers. The system sends this 512-dimensional one-dimensional feature vector into a fully connected classifier network via a data bus. The fully connected network consists of two cascaded matrix mapping layers. The first fully connected layer contains 256 physical neuron nodes, uses the ReLU function to increase non-linear isolation, and is connected in series with a Dropout layer with a forced inactivation rate of 0.5 to physically disconnect some neuron connections to prevent overfitting. After dimensionality reduction, the data enters the last fully connected layer at the end of the network. The number of physical neuron nodes in this output layer is equal to the number of wear level categories preset in the industrial field. In this embodiment, four physical levels are set: no wear, light wear, moderate wear, and heavy wear, so the number of output nodes is 4. This last layer outputs an original one-dimensional vector containing 4 real-valued scores. The system applies the Softmax probability distribution normalization function to these 4 scores.

[0050] For example, assuming the four original real-valued scores from the last layer output are [1.5, 0.2, 4, -1], after exponential normalization using the Softmax function, these four real numbers are converted into a probability distribution vector whose sum of internal elements is absolutely equal to 1, resulting in [0.07, 0.02, 0.9, 0.01]. This result indicates that the probability of the pipe wall image belonging to a normal, wear-free state is 7%, the probability of light wear is 2%, the probability of moderate wear is 90%, and the probability of heavy wear is 1%. Finally, the system's central processing unit executes the Argmax addressing instruction to extract the index of the element with the largest value from the above probability distribution vector. Based on this index, the system directly outputs a unique detection and diagnostic result to the terminal display of the industrial control system: the physical surface of this wear-resistant pipe section exhibits a "moderate wear" level.

[0051] Reference Figure 2Traditional ReLU functions directly set the weak pre-activation values ​​of the negative half-axis to zero, resulting in complete loss of features. Traditional LeakyReLU, while introducing a slope, has a globally fixed slope and cannot distinguish between signal and noise. The nonlinear activation processing provided by this invention can dynamically adjust the negative slope based on local structure factors: in regions with minor wear, a larger negative slope is generated, thereby preserving and amplifying the weak wear features with high fidelity; in smooth, wear-free background regions, the negative slope converges to the basic leakage coefficient, achieving strong suppression of background noise.

[0052] This invention also discloses an image processing-based wear-resistant pipe surface wear detection system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the image processing-based wear-resistant pipe surface wear detection method according to this invention.

[0053] The aforementioned image processing-based wear-resistant pipe surface wear detection system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0054] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for detecting wear on the surface of wear-resistant pipes based on image processing, characterized in that, include: The image to be detected on the surface of the wear-resistant pipe is acquired and then input into a multi-layer convolutional feature extraction network for hierarchical feature extraction to obtain feature maps of each layer. After at least one convolutional layer of the multi-layer convolutional feature extraction network, the pre-activation value of each neuron in the feature map output by the convolutional layer is subjected to non-linear activation processing. Specifically, when the pre-activation value is greater than zero, the output of the neuron is the pre-activation value. When the preactivation value is less than or equal to zero, the neuron's output is the product of the preactivation value and the compound negative slope coefficient; The composite negative slope coefficient is calculated based on spatial adjustment parameters, local structure factors, and a basic leakage coefficient; the spatial adjustment parameters are parameter values ​​in the trainable spatial sensitivity matrix corresponding to the neuron's position; the local structure factors are local feature dispersion indices calculated based on the local neighborhood features centered on the neuron in the current layer feature map; and the basic leakage coefficient is a preset positive hyperparameter. Based on the feature representation obtained after nonlinear activation processing, the wear characteristics of the wear-resistant tube surface are determined, including wear area or wear level.

2. The method for detecting wear on the surface of wear-resistant pipes based on image processing according to claim 1, characterized in that, The configuration method of the trainable spatial sensitivity matrix includes: initializing all elements in the spatial sensitivity matrix to preset positive constants before model training begins; and performing gradient calculation and parameter update on the spatial sensitivity matrix through backpropagation algorithm combined with a set loss function during model training.

3. The image processing-based method for detecting wear on the surface of a wear-resistant pipe according to claim 1, characterized in that, The acquisition of the local structure factor includes: calculating the local variance or local standard deviation of the current layer feature map using a neighborhood window of a preset size; generating a local discreteness feature map of the same size as the feature map in real time; obtaining the feature value corresponding to the neuron coordinates based on the real-time calculated local discreteness feature map; and normalizing the feature value to the interval between zero and one using a linear mapping function to obtain the local structure factor.

4. The method for detecting wear on the surface of wear-resistant pipes based on image processing according to claim 1, characterized in that, The basic leakage coefficient is a constant with a value between zero and one.

5. The method for detecting wear on the surface of wear-resistant pipes based on image processing according to claim 1, characterized in that, The process of obtaining the image to be detected on the surface of the wear-resistant pipe includes: uniformly illuminating the surface of the wear-resistant pipe to be detected using polarized light or diffuse reflection light source; acquiring the original image of the surface of the wear-resistant pipe using image acquisition hardware; adjusting the pixel size of the original image to match the preset input dimension of the multi-layer convolutional feature extraction network; and performing numerical normalization operations on the mean and standard deviation of the pixel matrix of the original image to obtain the image to be detected.

6. The method for detecting wear on the surface of wear-resistant pipes based on image processing according to claim 1, characterized in that, When the wear feature is a wear region, determining the wear feature of the wear-resistant pipe surface includes: inputting the feature representation into the decoder feature reconstruction network; by performing at least one upsampling or deconvolution calculation operation, gradually restoring the spatial feature size of the feature map to a predetermined size consistent with the image to be detected, generating a pixel-level feature prediction map; applying a nonlinear probability mapping function to each pixel feature value in the feature prediction map to calculate the probability calibration value of each independent pixel belonging to the wear region; comparing the probability calibration value of each independent pixel with a preset judgment threshold to determine whether each independent pixel belongs to the wear region, thereby generating a binary mask map composed of background pixels and wear pixels to obtain the wear region on the wear-resistant pipe surface.

7. The image processing-based method for detecting wear on the surface of a wear-resistant pipe according to claim 6, characterized in that, The computation of the decoder feature reconstruction network includes: after each upsampling or deconvolution operation, extracting shallow feature maps with the same physical size from the multi-layer convolutional feature extraction network; and merging the shallow feature maps with the new feature maps obtained after the upsampling or deconvolution operation in the channel dimension to form a fused map with multi-scale features.

8. The method for detecting wear on the surface of wear-resistant pipes based on image processing according to claim 1, characterized in that, When the wear feature is a wear level, determining the wear feature of the wear-resistant pipe surface includes: performing a global spatial dimension pooling operation on the feature representation to convert the multi-channel two-dimensional feature matrix into a one-dimensional feature vector of fixed length; inputting the converted one-dimensional feature vector into a feature mapping structure containing at least one fully connected layer for feature space transformation and dimension compression adjustment; setting the number of output computing nodes of the last fully connected layer to be equal to the preset number of wear level classifications; applying a probability distribution normalization mapping function to the feature vector output by the last fully connected layer to calculate the probability distribution values ​​of the current image to be detected belonging to each preset wear level, and extracting the category label corresponding to the maximum probability distribution value as the final determined wear level.

9. The method for detecting wear on the surface of a wear-resistant pipe based on image processing according to claim 1, characterized in that, The acquisition of the composite negative slope coefficient specifically includes: multiplying the local structure factor with the spatial adjustment parameter at the corresponding position to form a dynamic adjustment term, performing an additive bias operation on the dynamic adjustment term and the preset basic leakage coefficient to obtain a preliminary coefficient, and then restricting the preliminary coefficient to the range of zero to one by a truncation function to obtain the final composite negative slope coefficient.

10. A wear-resistant pipe surface wear detection system based on image processing, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the image processing-based wear detection method for wear-resistant pipe surfaces according to any one of claims 1-9.

Images