An image processing method and system for beam crack depth based on MLP optimization

Abstract

The present invention discloses an image processing method and system for beam crack depth based on MLP optimization, specifically relating to the field of neural network-based image data processing. The method involves converting raw image data into a uniformly spatially distributed image set through noise suppression, contrast remapping, and size normalization, and generating a structurally labeled image set with crack region labels through structural calibration and spatial mask construction. By constructing a spatial expression vector driven by image state evolution and introducing a variant MLP model with channel decoupling and path regulation, the method achieves crack region depth prediction processing driven by image structural changes, thereby improving the model's adaptability and prediction stability to complex image structures and addressing issues such as inconsistent image semantics, lack of spatial modeling, and insufficient depth prediction accuracy in existing methods.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing based on neural networks, and more specifically, to an image processing method and system for beam crack depth based on MLP optimization. Background Art

[0002] In existing crack detection and depth prediction technologies, although traditional image data processing methods can extract crack contour information to a certain extent, they generally suffer from problems such as non-uniform image spatial distribution, rough semantic structure annotation, and limited subsequent depth modeling and prediction accuracy.

[0003] Especially when faced with engineering images with complex structures, uneven lighting, or severe texture interference, it is impossible to form a stable and generalizable feature representation;

[0004] On the other hand, current neural network-based crack analysis methods mostly rely on shallow representations or end-to-end structures, lacking systematic modeling of the internal state evolution process of the image. This makes it difficult to implement a prediction mechanism dominated by the evolution laws of the local structure of the image. As a result, the predicted depth is sensitive to noise perturbations and lacks spatial guidance capabilities, ultimately affecting the model's interpretability and generalization capabilities.

[0005] Therefore, the technical problem to be solved by this solution is: how to construct a crack depth prediction model that integrates the image processing evolution mechanism and the neural network structure variation mechanism to realize image data processing for crack areas. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an image processing method and system for beam crack depth based on MLP optimization. By constructing a spatial expression vector driven by image state evolution and introducing a variant MLP model with channel decoupling and path control structure, the depth prediction processing of the crack area dominated by image structure changes is realized, thereby improving the model's adaptability and prediction stability to complex image structures, and solving the problems of existing methods such as inconsistent image semantics, lack of spatial modeling, and insufficient depth prediction accuracy.

[0007] To achieve the above-mentioned object, the present invention provides the following technical solutions: an image processing method for beam crack depth based on MLP optimization, comprising acquiring original image data through an image acquisition device;

[0008] S1. The original image data is converted into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization. Then, a set of structurally annotated images with crack area labels is generated through structural calibration and spatial mask construction.

[0009] S2. Construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation;

[0010] S3. Map the image processing evolution table to the crack area according to the spatial position to construct a set of prediction probability distribution vectors. Then, through entropy density calculation and direction field construction, an expression vector with a spatial guidance mechanism is generated as the input feature of the variant MLP model.

[0011] S4. Construct an image data processing structure based on a neural network: Divide the hidden layer of the MLP model into three layers, and form a variant MLP model by introducing an evolutionary grid layer and a path control mechanism;

[0012] S5. The expression vector with the spatial guidance mechanism is sequentially input into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of the structural channel decoupling, and the third hidden layer of the path control. The state propagation, feature superposition, and path control weight control operations are performed in sequence, and finally the depth prediction value of the corresponding crack area is generated in the output layer.

[0013] In a preferred embodiment, in S1, original image data is acquired by an image acquisition device, and an image dataset including a pixel matrix and spatial position information is formed based on the original image data. The image dataset is subjected to a Gaussian filter to perform noise suppression processing to form a background interference suppressed image set.

[0014] The background interference suppressed image set is processed through grayscale mapping and contrast remapping to generate a crack boundary information enhanced image set. The crack boundary information enhanced image set is spatially cropped and size normalized to form an image set with uniform spatial distribution. The image set is then converted into a structure-annotated image set through regional structure calibration and spatial mask construction operations. The structure-annotated image set includes crack region block labels and spatial structure references.

[0015] In a preferred embodiment, in S2, a set of structurally annotated images is obtained, and an initial spatial state tensor is generated by performing a feature mapping operation on the set of structurally annotated images. The initial spatial state tensor includes a local edge gradient, a grayscale value, and a structural label mapping value. The initial spatial state tensor is then state-advanced by a diffusion-driven propagation mechanism to generate a first evolutionary state map of the crack.

[0016] The first evolutionary state map is reconstructed by performing path continuity reconstruction through a directional gradient reconstruction operation to form a crack direction state map. The crack direction state map is combined with a material property distribution function to perform structural prior effect mapping to generate a second evolutionary state map under the influence of material structure.

[0017] The second evolutionary state map constructs a multi-step state evolution trajectory through a sequence stepping mechanism to form a state evolution sequence tensor; the state evolution sequence tensor is compressed and aggregated to construct a deep space state expression vector as an image processing evolution table.

[0018] In a preferred embodiment, in S3, an image processing evolution table is obtained, and after the image semantic state embedding features in the image processing evolution table are mapped to the corresponding crack area according to the spatial position, continuous value prediction distribution fitting is performed respectively to construct a set of prediction probability distribution vectors of the crack area; the set of prediction probability distribution vectors is processed by local entropy density calculation to form a regional prediction entropy tensor, and the regional prediction entropy tensor is used to identify the degree of information uncertainty of each image area;

[0019] The regional prediction entropy tensor is used to form an entropy change direction field through gradient calculation, and the entropy change direction field is used to describe the flow direction of information uncertainty in space;

[0020] The entropy change direction field and the image processing evolution table are fused through path vector processing to construct a path guidance feature vector. The path guidance feature vector is subjected to a step-by-step feedback superposition operation to generate an expression vector with a spatial guidance mechanism. The expression vector is used as the input feature of the variant MLP model.

[0021] In a preferred embodiment, S4 further includes an MLP model and a variant MLP model; the MLP model includes an input layer, a hidden layer, and an output layer; the hidden layer of the MLP model is divided into a first hidden layer, a second hidden layer, and a third hidden layer; a variant MLP model is constructed based on the MLP model, the variant MLP model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer;

[0022] The input layer of the variant MLP model is used to receive the expression vector with a spatial guidance mechanism and complete the initial projection mapping;

[0023] The first hidden layer of the mutated MLP model is replaced by an evolving lattice field layer, which performs state construction by simulating the spatial propagation process of structural states in the image.

[0024] The second hidden layer of the variant MLP model is divided into a crack morphology channel, a material response channel, and a path evolution channel through structural channel decoupling, and performs independent feature mapping respectively;

[0025] The third hidden layer of the variant MLP model retains the fully connected structure and introduces a path control weight adjustment operation;

[0026] The output layer of the variant MLP model is a linear mapping structure, which is used to output the continuous prediction value of the crack depth;

[0027] The variant MLP model is jointly constructed by a structural state propagation mechanism and a path information regulation mechanism, and is used to establish a mapping relationship from image features to crack depths.

[0028] In a preferred embodiment, in S5, the expression vector with the spatial guidance mechanism is input into the input layer of the variant MLP model and is transformed into an input projection tensor after an initial weight transformation;

[0029] The input projection tensor is input to the first hidden layer, and the state propagation processing is performed through the evolved lattice mechanism of the evolved lattice layer, generating a state propagation representation tensor in the process of structural information diffusion and expression reconstruction;

[0030] The state propagation representation tensor undergoes a decoupling transformation through the structural channel and enters the crack morphology channel, material response channel, and path evolution channel to perform independent feature mapping. The feature tensors of the three channels are output respectively. The feature tensors of the three channels are fused through feature superposition to construct a composite feature tensor. The second hidden layer performs activation function nonlinear transformation and path weight redistribution processing on the composite feature tensor to generate a path-aware fusion vector.

[0031] The path-aware fusion vector is input into the third hidden layer, and the fully connected mapping and path control weight adjustment operations are performed to generate the final regression representation vector;

[0032] The final regression representation vector is input to the output layer, where the depth prediction value of the corresponding crack area is generated through linear mapping.

[0033] In a preferred embodiment, in S2, the structure annotation image set Through the feature mapping function Mapped to the initial spatial state tensor , Including local edge gradient, gray value and structure label mapping value; using diffusion driven propagation mechanism to Execute the state advancement operation to simulate the crack evolution and expansion behavior and generate the first evolution state map ; Directed gradient reconstruction operation , perform path continuity recovery and generate crack direction state map ;Will Fusion Material Property Distribution Function and structural prior mapping fields , generating a second evolutionary state map under the influence of material structure ; Then through the sequence stepping mechanism Multi-step advancement to build state evolution sequence tensor ; will eventually Perform compression aggregation to generate deep spatial state expression vectors , as the image processing evolution table;

[0034] Perform feature mapping and state tensor generation:

[0035] ;

[0036] ;

[0037] ;

[0038] Diffusion-driven first-order state evolution:

[0039] ;

[0040] Perform directional gradient reconstruction operation:

[0041] ;

[0042] Expression of structural prior mapping effect:

[0043] ;

[0044] Expression for multi-step state advancement:

[0045] ;

[0046] Compression aggregation into deep expression vector:

[0047] ;

[0048] The structure annotation image collection Used to provide spatial identification and label information of the crack area; feature mapping function Used to 、 and The joint encoding is the initial spatial state tensor; is the gradient tensor of the image; image grayscale normalization function Used to convert grayscale images The pixel values ​​in are linearly mapped to interval; is the crack area label mapping function, which is an indicator function used to map the crack area label at each pixel. The position of the crack area is marked as belonging to the crack area, and the output of the crack area label mapping function is Binary response plot of ; is the initial spatial state tensor; image Represents the pixel grayscale matrix of the original input image; Representing an image The lower limit of all pixel grayscale values ​​in ; Representing an image The upper limit of all pixel grayscale values ​​in ; is a Boolean-valued function; Representing an image Pixels in Whether it belongs to the crack area set ;

[0049] The local diffusion coefficient tensor Indicates time Initial spatial state tensor function, Indicates time The initial spatial state tensor under , Used to characterize the information flow ability of each position in the image, The calculation form is: ,in is the diffusion inhibition factor, represents the square of the Euclidean two-norm, Express Perform directional gradient reconstruction operation, represents the spatial gradient operator, is an exponential function; is the global diffusion rate constant during the diffusion process; time For historical events, time is the current time; The termination diffusion time point of the first-level evolution process; Indicates time Approaching When , the final convergence value; is the first evolution state spectrum; the symbol · in the formula is the dot product;

[0050] The directional gradient reconstruction operation In the formula, it is expressed as the directional gradient reconstruction function; where * is the convolution operation, is the component of the spatial gradient operator in the main direction; is the component of the spatial gradient operator in the vertical direction; is a directional convolution kernel, which is used to highlight the main direction structural features of the crack. The calculation form is:

[0051] ;

[0052] The direction is the local main direction angle; For direction Vertical convolution kernel; is the crack direction state map;

[0053] in is the material property distribution function; the structural prior mapping field In the formula, it represents the structural prior mapping function; is the structural attribute modulation coefficient; is the second evolutionary state map;

[0054] in The second evolutionary state map In time The evolutionary state of It is the state evolution promotion function, which is used to perform the update operation of each state step. The calculation form is:

[0055] ;

[0056] in is the Laplace operator, the hyperbolic tangent function In the formula, represents the nonlinear saturation transformation; is the diffusion control coefficient; is the structural attribute modulation coefficient; is the time step of state evolution; is the state evolution sequence tensor, which records the state evolution sequence tensor from to All state evolution graphs of ; is the time series length of the second evolution stage;

[0057] in is the tensor compression aggregation function; is the activation function; is the global average pooling function; For time The weight coefficient of the step state in the aggregation; Evolution table vector for image processing.

[0058] In a preferred embodiment, in S3, it further includes:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] in The crack area Medium pixel The predicted probability distribution value of ; For the region Normalization factor on ; Pixel points in the image processing evolution table Status characteristics; Pixel The predicted target value mapping; The center pixel of the crack With pixels The Euclidean distance between is the exponential scaling factor of the forecast difference; is the distance adjustment coefficient;

[0066] in Indicates the The set of all pixels in the crack area; Pixel The image gradient vector at ; Represents pixel points and The square of the second norm of the image gradient difference; Pixel The information entropy direction field vector; Pixel The divergence of the information entropy direction field at ; is the exponential control factor, defined as: ; Indicates the angle; is the crack area label mapping function, if the pixel Belong to the crack area, then ,otherwise ; is the crack density function, defined as: , Represents pixel points Neighborhood window of Represents pixel points A pixel in the neighborhood of ; Represents pixel points Crack area label mapping function;

[0067] in The crack area The predicted entropy density value of ; It is the minimum protection constant in entropy calculation, which is used to prevent the logarithmic term from going to infinity; is the direction vector of the regional entropy gradient; is the symbol of partial derivative; is the horizontal coordinate axis of the image in the two-dimensional plane, is the vertical coordinate axis of the image in the two-dimensional plane;

[0068] in Crack area in the image processing evolution table The original path feature vector of ; is the path history response operator; Represents the element-wise product between tensors; is the fusion operator; The crack area The guided path vector of

[0069] in The crack area The spatial guide expression vector of Finally, it serves as the input of the variant MLP model; For the feedback superposition Step adjustment weight of step; is the preset feedback step constant; is the activation function; is the step response function; is the step index of the current jump; Indicates the total number of steps of skip feedback superposition.

[0070] In a preferred embodiment, a variant MLP model is constructed in S4 and S5:

[0071] Input layer:

[0072] ;

[0073] First hidden layer:

[0074] ;

[0075] Second hidden layer:

[0076] ;

[0077] in:

[0078] ;

[0079] Third hidden layer:

[0080] ;

[0081] Output layer:

[0082] ;

[0083] in is the input mapping function; is the input feature vector after projection; is the evolution lattice function; is the weight mapping matrix of the input layer; is the bias vector of the input layer; The crack area The spatial neighborhood set of ; From the crack area To the crack area Propagation factor of crack area Indicates the current crack area In, with Indexes of adjacent pixels or regions that have spatial or path connectivity; is the grid field kernel function; The state propagation representation tensor output by the evolutionary grid field layer;

[0084] They are the feature extraction mapping functions for the crack morphology channel, material response channel, and path evolution channel respectively; Independent feature tensors output for three types of channels; is the composite feature tensor after channel fusion; is the path regulation mapping matrix of the third hidden layer; represents the information guided weight function; Indicates crack area superior About Pixels gradient; is the activation control function of the path regulation layer; is the weight mapping matrix of the path control layer; is the path-aware fusion vector output by the third hidden layer; The final output crack area The depth prediction value of is the weight transpose vector of the output layer; is the bias vector of the output layer.

[0085] An image processing system for beam crack depth based on MLP optimization, including a labeling module, an evolution module, a guided expression module, a mutation module, and a prediction output module;

[0086] The annotation module is used to convert the original image data into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization, and generate a set of structurally annotated images with crack area labels through structural calibration and spatial mask construction;

[0087] The evolution module is used to construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation;

[0088] The guided expression module is used to map the image processing evolution table to the crack area according to the spatial position and construct a set of prediction probability distribution vectors. It also generates an expression vector with a spatial guidance mechanism through entropy density calculation and direction field construction, which serves as the input feature of the variant MLP model.

[0089] The mutation module is used to construct an image data processing structure based on a neural network: the hidden layer of the MLP model is divided into three layers, and a mutation MLP model is formed by introducing an evolutionary grid layer and a path control mechanism;

[0090] The prediction output module is used to input the expression vector with a spatial guidance mechanism into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of structural channel decoupling, and the third hidden layer of path regulation in sequence, and then perform state propagation, feature superposition, and path control weight regulation operations, and finally generate the depth prediction value of the corresponding crack area in the output layer.

[0091] The technical effects and advantages of the present invention are as follows:

[0092] 1. This solution systematically addresses the problems of chaotic spatial distribution, rough structural annotation, and weak deep modeling capabilities in traditional image processing by building a prediction model that integrates the image processing evolution mechanism with the neural network structural variation mechanism. This improves the stability and prediction accuracy of image processing in crack areas.

[0093] 2. By performing noise suppression, contrast remapping, and unified spatial normalization on the original image, combined with structure calibration and mask construction operations, the clarity and semantic labeling of crack areas in the image are enhanced, laying a high-quality input foundation for subsequent modeling;

[0094] 3. A three-stage state evolution mechanism is used to integrate image feature diffusion, directional reconstruction, and material prior expression to construct a cross-scale, highly correlated spatial state sequence expression, addressing the shortcomings of traditional expressions in modeling structural evolution processes.

[0095] 4. By constructing the local entropy density and direction field of the predicted probability vector, a perceptible path guidance feature of the information flow is formed. Combined with the step-by-step feedback mechanism to form an expression vector, the spatial adaptability and expression guidance capabilities of the model in complex structural scenarios are effectively improved;

[0096] 5. Construct a variant MLP network with an evolutionary grid, channel decoupling, and path regulation structure, combining layer-by-layer feature flow control and structural state propagation mechanisms to enhance the neural network's ability to respond to image evolution and crack depth regression performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 The figure is a flow chart of the method steps of the present invention.

[0098] Figure 2 Schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0099] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0100] Refer to the instruction manual Figure 1-2 , an image processing method for beam crack depth based on MLP optimization according to an embodiment of the present invention includes acquiring original image data through an image acquisition device;

[0101] S1. The original image data is converted into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization. Then, a set of structurally annotated images with crack area labels is generated through structural calibration and spatial mask construction.

[0102] S2. Construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation;

[0103] S3. Map the image processing evolution table to the crack area according to the spatial position to construct a set of prediction probability distribution vectors. Then, through entropy density calculation and direction field construction, an expression vector with a spatial guidance mechanism is generated as the input feature of the variant MLP model.

[0104] S4. Construct an image data processing structure based on a neural network: Divide the hidden layer of the MLP model into three layers, and form a variant MLP model by introducing an evolutionary grid layer and a path control mechanism;

[0105] S5. The expression vector with the spatial guidance mechanism is sequentially input into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of the structural channel decoupling, and the third hidden layer of the path control. The state propagation, feature superposition, and path control weight control operations are performed in sequence, and finally the depth prediction value of the corresponding crack area is generated in the output layer.

[0106] In S1, original image data is acquired through an image acquisition device, and an image data set including a pixel matrix and spatial position information is formed based on the original image data. The image data set is subjected to a noise suppression process by Gaussian filtering to form a background interference suppressed image set.

[0107] The background interference suppressed image set is processed through grayscale mapping and contrast remapping to generate a crack boundary information enhanced image set. The crack boundary information enhanced image set is spatially cropped and size normalized to form an image set with uniform spatial distribution. The image set is then converted into a structure-annotated image set through regional structure calibration and spatial mask construction operations. The structure-annotated image set includes crack region block labels and spatial structure references.

[0108] In S2, a set of structurally annotated images is obtained and subjected to a feature mapping operation to generate an initial spatial state tensor. The initial spatial state tensor includes local edge gradients, grayscale values, and structural label mapping values. The initial spatial state tensor is then state-advanced through a diffusion-driven propagation mechanism to generate a first evolutionary state map of the crack. This stage is defined as the first-level evolution process.

[0109] The first evolutionary state map is reconstructed through path continuity reconstruction using a directional gradient reconstruction operation to form a crack directional state map. The crack directional state map is combined with the material property distribution function to perform structural prior mapping to generate a second evolutionary state map under the influence of the material structure. This stage is defined as secondary evolution processing.

[0110] The second evolutionary state map constructs a multi-step state evolution trajectory through a sequence stepping mechanism to form a state evolution sequence tensor; the state evolution sequence tensor is compressed and aggregated to construct a deep spatial state expression vector as the image processing evolution table. This stage is defined as three-level evolution processing.

[0111] In S3, an image processing evolution table is obtained. After the image semantic state embedding features in the image processing evolution table are mapped to the corresponding crack areas according to the spatial positions, continuous value prediction distribution fitting is performed respectively to construct a set of prediction probability distribution vectors of the crack areas; the set of prediction probability distribution vectors is processed by local entropy density calculation to form a regional prediction entropy tensor, and the regional prediction entropy tensor is used to identify the degree of information uncertainty of each image area;

[0112] The regional prediction entropy tensor is used to form an entropy change direction field through gradient calculation, and the entropy change direction field is used to describe the flow direction of information uncertainty in space;

[0113] The entropy change direction field and the image processing evolution table are fused through path vector processing to construct a path guidance feature vector. The path guidance feature vector is subjected to a step-by-step feedback superposition operation to generate an expression vector with a spatial guidance mechanism. The expression vector is used as the input feature of the variant MLP model.

[0114] S4 also includes an MLP model and a variant MLP model; the MLP model includes an input layer, a hidden layer, and an output layer; the hidden layer of the MLP model is divided into a first hidden layer, a second hidden layer, and a third hidden layer; a variant MLP model is constructed based on the MLP model, the variant MLP model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer;

[0115] The input layer of the variant MLP model is used to receive the expression vector with a spatial guidance mechanism and complete the initial projection mapping;

[0116] The first hidden layer of the mutated MLP model is replaced by an evolving lattice field layer, which performs state construction by simulating the spatial propagation process of structural states in the image.

[0117] The second hidden layer of the variant MLP model is divided into a crack morphology channel, a material response channel, and a path evolution channel through structural channel decoupling, and performs independent feature mapping respectively;

[0118] The third hidden layer of the variant MLP model retains the fully connected structure and introduces a path control weight adjustment operation;

[0119] The output layer of the variant MLP model is a linear mapping structure, which is used to output the continuous prediction value of the crack depth;

[0120] The variant MLP model is jointly constructed by a structural state propagation mechanism and a path information regulation mechanism, and is used to establish a mapping relationship from image features to crack depths.

[0121] In S5, the expression vector with spatial guidance mechanism is input into the input layer of the variant MLP model and forms the input projection tensor after initial weight transformation;

[0122] The input projection tensor is input to the first hidden layer, and the state propagation processing is performed through the evolved lattice mechanism of the evolved lattice layer, generating a state propagation representation tensor in the process of structural information diffusion and expression reconstruction;

[0123] The state propagation representation tensor undergoes a decoupling transformation through the structural channel and enters the crack morphology channel, material response channel, and path evolution channel to perform independent feature mapping. The feature tensors of the three channels are output respectively. The feature tensors of the three channels are fused through feature superposition to construct a composite feature tensor. The second hidden layer performs activation function nonlinear transformation and path weight redistribution processing on the composite feature tensor to generate a path-aware fusion vector.

[0124] The path-aware fusion vector is input into the third hidden layer, and the fully connected mapping and path control weight adjustment operations are performed to generate the final regression representation vector;

[0125] The final regression representation vector is input to the output layer, where the depth prediction value of the corresponding crack area is generated through linear mapping.

[0126] It should be noted that in the formula structure involved in this solution, dimensionless terms can serve as proportionality or structural adjustment factors. When combined with quantities with units, they only play a numerical scaling role and do not introduce new physical dimensions. Therefore, they will not change or confuse the overall unit system of expression. This combination of "dimensionless terms and units" can be understood as a composite structural expression commonly used in mathematical and physical modeling, conforming to the principle of dimensional consistency and having a clear physical interpretation basis.

[0127] Secondly, in the formula structure of this scheme, if multiple variables with different physical units are involved, including but not limited to time, mass or energy variables, their joint appearance is to express the collaborative modeling relationship of multiple physical mechanisms. Each variable can be formed into a unified structure through function mapping, ratio combination or normalization adjustment. The units and meanings are clear, and the overall expression conforms to the principle of dimensional consistency and the common formula of engineering modeling.

[0128] Any constants, weights, adjustment factors, threshold parameters, and proportional coefficients involved in this solution are all adjustable control parameters for different application environments. Their values ​​depend on the target device configuration, data input characteristics, and performance optimization goals. During the implementation phase, they are set within a reasonable range through model verification, performance constraints, or engineering calibration. Although these parameters do not have preset unique values, they have clear adjustment logic and calculation paths and are part of the deterministic setting process in engineering implementation. The purpose of such setting is to ensure that the solution is both universally adaptable, reproducible, and operable, without affecting its technical clarity and feasibility.

[0129] In S2, the structure annotation image collection Through the feature mapping function Mapped to the initial spatial state tensor , Including local edge gradient, gray value and structure label mapping value; using diffusion driven propagation mechanism to Execute the state advancement operation to simulate the crack evolution and expansion behavior and generate the first evolution state map (first-level evolution); Directed gradient reconstruction operation , perform path continuity recovery and generate crack direction state map ;Will Fusion Material Property Distribution Function and structural prior mapping fields , generating a second evolutionary state map under the influence of material structure (Secondary evolution); then through the sequence stepping mechanism Multi-step advancement to build state evolution sequence tensor (Level 3 Evolution); eventually Perform compression aggregation to generate deep spatial state expression vectors , as the image processing evolution table;

[0130] Perform feature mapping and state tensor generation:

[0131] ;

[0132] ;

[0133] ;

[0134] Diffusion-driven first-order state evolution (first-order evolution):

[0135] ;

[0136] Perform directional gradient reconstruction operation:

[0137] ;

[0138] Expression of structural prior mapping (secondary evolution):

[0139] ;

[0140] Expression for multi-step state advancement (three-level evolution):

[0141] ;

[0142] Compression aggregation into deep expression vector:

[0143] ;

[0144] The structure annotation image collection Used to provide spatial identification and label information of crack areas. The structure annotation image set represents the image data set after structural calibration and spatial mask processing, including but not limited to: structural calibration performs Sobel operator and local gradient extreme value extraction operations on the background interference suppression image set to identify the boundary lines with prominent pixel gradient changes, and combines regional connectivity analysis to construct a structural label map of the crack area. Spatial mask processing constructs a binary spatial mask based on the structural label map, and performs pixel-by-pixel product operation with the original image set through the mask matrix to shield the interference information of the non-crack area and strengthen the structural boundary expression; feature mapping function Used to 、 and The joint encoding is the initial spatial state tensor; The gradient tensor of the image is used to represent the first-order change of the pixel gray value in the spatial dimension, reflecting the strength and direction of the local edge; the image gray normalization function Used to convert grayscale images The pixel values ​​in are linearly mapped to Interval, used to standardize brightness changes and enhance the stability of contrast response to structure. The image grayscale normalization function includes linearly converting the original grayscale value of each pixel in the image according to a certain preset ratio to map it to a set target grayscale range (such as 0 to 1 or 0 to 255). This process is usually based on the minimum and maximum grayscale values ​​of the entire image for normalization calculation, thereby unifying the image brightness scale and enhancing the stability of subsequent processing; is the crack area label mapping function, which is an indicator function used to map the crack area label at each pixel. The position of the crack area is marked as belonging to the crack area, and the output of the crack area label mapping function is Binary response plot of ; is the initial spatial state tensor, which is the output of the feature mapping function and is used as the starting input for state diffusion and evolution; image Represents the pixel grayscale matrix of the original input image, image Is the input object of the normalization operation; Representing an image The lower limit of all pixel grayscale values ​​in ; Representing an image The upper limit of all pixel grayscale values ​​in ; is a Boolean function used to give the image Label the crack areas; Representing an image Pixels in Whether it belongs to the crack area set ;

[0145] The local diffusion coefficient tensor Indicates time Initial spatial state tensor function, Indicates time The initial spatial state tensor under , Used to characterize the information flow ability of each position in the image, The calculation form is: ,in is the diffusion inhibition factor, represents the square of the Euclidean two-norm, Express Perform directional gradient reconstruction operation, Represents the spatial gradient operator, which is used to calculate the rate of change of the tensor field in the spatial dimension. is an exponential function; the size of the diffusion suppression factor is determined according to the severity of the structural change in the image. When the structural change in a certain area is more severe (that is, the more obvious the edge), the gradient value at the corresponding position is larger, and the diffusion suppression factor becomes smaller, indicating that diffusion should be strongly suppressed at this position to avoid blurred edges. On the contrary, if a certain area changes slowly and the gradient value is small, the diffusion suppression factor is close to 1, indicating that diffusion is allowed to achieve smooth propagation. The parameter that adjusts this diffusion suppression factor is called the adjustment coefficient, which determines the sensitivity of the system to the structural boundary. In general, the value of the diffusion suppression factor reflects whether the position is suitable for diffusion. The closer it is to 1, the more suitable it is for diffusion, and the closer it is to 0, the less suitable it is for diffusion. is the global diffusion rate constant during the diffusion process. In the diffusion-driven propagation mechanism, express The overall diffusion capacity in the spatial domain needs to be set according to the sensitivity of the crack boundary in the structural image and the demand for the diffusion range: if the crack area boundary is clear and the local changes are drastic, a smaller value should be set. (include ), in order to prevent the diffusion process from crossing the real structure boundary, if the crack area in the image has blurred edges or there is a low contrast area, it can be appropriately increased (include ), to enhance the continuity of the state in the fuzzy area; finally The setting of should be based on the average gradient response distribution of the fracture structure in the training set, combined with the gradient intensity threshold and spatial smoothness requirements, and the optimal parameter selection should be made through cross-validation or based on the information entropy decrease rate; time For historical events, time is the current time; The termination diffusion time point of the first-level evolution process; Indicates time Approaching When , the final convergence value; is the first evolutionary state map, which represents the After applying the diffusion-driven propagation mechanism, The stable spatial state reached at the moment represents the expansion form of the initial crack structure; the symbol · in the formula is the dot product;

[0146] The directional gradient reconstruction operation In the formula, it is represented by the directional gradient reconstruction function, which is used to perform gradient weighted convolution operations along the main direction of the crack and its perpendicular direction to reconstruct the path continuity; where * is the convolution operation, is the component of the spatial gradient operator in the main direction, which is used to extract the rate of change information along the main axis of the crack or structure in the image; It is the component of the spatial gradient operator in the vertical direction. The component of the spatial gradient operator in the vertical direction is used to capture the edge or texture changes perpendicular to the main structure direction in the image; is a directional convolution kernel, which is used to highlight the main direction structural features of the crack. The calculation form is:

[0147] ;

[0148] The direction is the local main direction angle; For direction The vertical convolution kernel, Used to obtain structural change characteristics in orthogonal directions; The fracture direction state map is obtained by reconstructing the directional gradient and is used to reflect the coherence and topological direction of the fracture path.

[0149] in is the material property distribution function, which is used to define the position in the image coordinate space The spatial variation of local material properties is represented by the above. Local material properties include but are not limited to mechanical impedance, microstructure texture strength, etc. The calculation form is but not limited to: ,in is the density field or anisotropy field, is the weighting coefficient; Indicates that the image is The pixel intensity value at the point is used to describe the grayscale or brightness information of the image at that point; the structural prior mapping field In the formula, it represents the structural prior mapping function, which combines the directional map with the material properties to correct the structural propagation path; is the structural attribute modulation coefficient, which is used to The value of the structural attribute modulation coefficient includes adjustment according to the clarity of the crack boundary and the complexity of the material texture in the training image, including but not limited to selection within the interval [0.1, 1.0]: if the crack boundary is fuzzy and the material texture is complex, the value can be increased. , to increase the strength of structural response. If the structural features are too sensitive or boundary noise occurs, you can reduce , to prevent overfitting error diffusion; The second evolutionary state map integrates the material distribution factors and the crack path structure, and is the basic state map for the final multi-step sequence evolution.

[0150] in The second evolutionary state map In time The evolution state of is a single time frame in the sequence tensor; It is the state evolution promotion function, which is used to perform the update operation of each state step. The calculation form is:

[0151] ;

[0152] in is the Laplace operator, the hyperbolic tangent function In the formula, represents the nonlinear saturation transformation; is the diffusion control coefficient, which is used to adjust The guided global diffusion intensity and the diffusion control coefficient are set according to the degree of structural change in the crack image. The smoother the structural boundary, the larger η is, which is used to enhance the continuous diffusion ability. is the structural attribute modulation coefficient, which is used to measure the nonlinear saturation mapping term The influence of state regulation is defined by the magnitude of the image grayscale change. The stronger the boundary contrast, the larger the λ is, so as to enhance the ability to maintain the structural mutation area. is the time step of state evolution, which is used to control the discrete iteration interval of continuous evolution; is the state evolution sequence tensor, which records the state evolution sequence tensor from to All state evolution maps are used to comprehensively express the dynamic expansion trajectory of cracks under the influence of structure; is the time series length of the second evolution stage;

[0153] in It is a tensor compression aggregation function, which is used to reduce the dimension of the state evolution sequence into a single deep expression vector. Its internal operations include feature transformation, activation and weighted superposition; The activation function includes ReLU or GELU, which is used to nonlinearly enhance the sparsity of state features; The global average pooling function is used to average each feature channel in the spatial dimension to retain high-level semantic compression information; For time The weight coefficient of the step state in the aggregation, time The weight coefficient of the state of the step in the aggregation is used to adjust the contribution of each time frame to the final expression vector. In practical applications, it can be given by the attention mechanism or prior weight. By constructing an attention module based on position encoding and channel weighting, the time frame in the state evolution sequence can be The representation vector of the time step is weighted and calculated to generate a dynamic weight coefficient for compression aggregation. At the same time, the static weight ratio of different time steps can be set by introducing a priori characteristic functions based on the length of the crack area, morphological complexity or material properties, thereby achieving effective integration and practical expression of evolution information in the image processing evolution table. The image processing evolution table vector is the final image feature expression obtained by compressing the aggregated state sequence tensor, and is used as the input feature of the subsequent depth prediction model.

[0154] In S3, it also includes:

[0155] ;

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] in The crack area Medium pixel The predicted probability distribution value of indicates the possibility that the location is a crack area; For the region The normalization factor on is a valid probability distribution; Pixel points in the image processing evolution table Status characteristics; Pixel The predicted target value mapping; The center pixel of the crack With pixels The Euclidean distance between is the exponential scaling factor of the prediction difference, which is used to control the sensitivity of the distribution; is the distance adjustment coefficient, which is used to enhance the effect of position difference on probability;

[0162] in Indicates the The set of all pixels in the crack area, Determined by the corresponding label in the structure annotation image set; Where the crack center pixel Serves as a benchmark reference point for directionality and gradient similarity; Pixel The image gradient vector at , which represents the image at and Grayscale change rate of direction; Represents pixel points and The square of the image gradient difference is used to measure the similarity of edge direction and texture structure; Pixel The information entropy direction field vector is used to represent the local flow trend of information uncertainty in the image; Pixel The divergence of the information entropy direction field at the point is used to reflect the degree of information aggregation or divergence at the point, affecting the intensity of information flow regulation; is the exponential control factor, defined as: , the exponential control factor represents the pixel The angle between the information entropy direction and the gradient direction is used to adjust the influence weight of the divergence in different directions; Indicates the angle; is the crack area label mapping function, if the pixel Belong to the crack area, then ,otherwise , the crack region label mapping function is used to highlight the importance of the crack region; is the crack density function, defined as: , Represents pixel points The neighborhood window of Used to count the number of crack points in a local area and reflect the concentration of regional cracks; Represents pixel points A pixel in the neighborhood of ; Represents pixel points Crack area label mapping function;

[0163] in The crack area The predicted entropy density value represents the uncertainty of information; It is the minimum protection constant in entropy calculation, which is used to prevent the logarithmic term from going to infinity; is the regional entropy gradient direction vector, which is used to represent the direction of uncertainty change; is the symbol of partial derivative; is the horizontal coordinate axis of the image in the two-dimensional plane, is the vertical coordinate axis of the image in the two-dimensional plane;

[0164] in Crack area in the image processing evolution table The original path feature vector of ; is the path history response operator, which is used to encode the history jump feedback mechanism; Represents the element-wise product between tensors; is the fusion operator, which is used to perform vector superposition and normalization; The crack area The guided path vector is the result of coupling between the spatial path and the entropy direction;

[0165] in The crack area The spatial guide expression vector of Finally, it serves as the input of the variant MLP model; For the feedback superposition The step adjustment weight of step, The value is obtained through The inner product square of the unit vector of each candidate direction is normalized to reflect the degree of dominance of the direction in the propagation of spatial information uncertainty. Specifically, The larger it is, the more significant the change in information entropy in that direction is, and the path guidance mechanism will be more inclined to focus features along that direction. Satisfy the normalization constraint, its value range is [0, 1], and the sum of the coefficients in all directions is 1; The preset feedback step constant is used to control the skip step size. As a tuning parameter for controlling the contribution of historical path features to current path guidance during the skip feedback operation, the feedback step constant is not derived from adaptive learning during model training but is manually set and determined through experimental tuning. It is a typical hyperparameter. In specific implementations, this constant is typically set within a limited range (e.g., 0.1 to 1.0) and optimized through cross-validation or grid search based on the model's performance on the validation set. This ensures path guidance consistency while suppressing spatial information lag caused by over-reliance on historical features. is the activation function, which includes ReLU or GELU; is the step response function, which is used to simulate the historical memory response of the path state; is the step index of the current jump; Indicates the total number of steps of skip feedback superposition.

[0166] Construct the mutation MLP model in S4 and S5:

[0167] Input layer:

[0168] ;

[0169] First hidden layer:

[0170] ;

[0171] Second hidden layer:

[0172] ;

[0173] in:

[0174] ;

[0175] Third hidden layer:

[0176] ;

[0177] Output layer:

[0178] ;

[0179] in The input mapping function is used to perform initial feature compression and activation transformation. The mapping of the input mapping function is used to control the input scale and enhance the nonlinear feature response; is the input feature vector after projection; The evolved lattice field function is used to define the information propagation and state update mechanism within the spatial neighborhood, and construct the structural state map by simulating the diffusion form of local differences; is the weight mapping matrix of the input layer, which is used to transform Mapping to a high-dimensional projection space; The bias vector of the input layer is used to adjust the feature distribution after linear mapping to improve the expression ability of nonlinear activation; The crack area The spatial neighborhood set of The spatial neighborhood set of is used for local grid evolution to control the propagation range and structural continuity of evolution information; From the crack area To the crack area The propagation factor, Depends on the position and structure differences, indicating the degree of influence of different grid points on the results during evolution; crack area Indicates the current crack area In, with Indexes of adjacent pixels or regions that have spatial or path connectivity; is the lattice field kernel function, which is used to calculate the propagation intensity under the change of spatial characteristics. In practical applications, the lattice field kernel function contains a nonlinear response function of the structure tensor difference metric; The state propagation representation tensor output by the evolutionary lattice field layer is used to describe the evolutionary reconstruction results of the structural information in the local space.

[0180] They are the feature extraction mapping functions of the crack morphology channel, material response channel, and path evolution channel, respectively, which are used to extract the feature expressions of structural changes, material disturbances, and path continuity in the image; are independent feature tensors output by the three types of channels, and the independent feature tensors output by the three types of channels correspond to the projection responses of the input state propagation tensor on different structural dimensions; It is the composite feature tensor after channel fusion, which is used to form the intermediate layer representation of cross-channel semantic connections; is the path regulation mapping matrix of the third hidden layer, which is used to perform high-dimensional weight adjustment on the composite feature tensor; represents the information-guided weight function, which is used to adjust the importance of the predicted probability gradient at different spatial positions according to the gradient information; Indicates crack area superior About Pixels The gradient, Used to describe the direction and rate of change of probability; is the activation control function of the path regulation layer, which is used to dynamically select feature significance and enhance the impact of path state changes on the final feature output; is the weight mapping matrix of the path control layer. The weight mapping matrix is ​​used to perform linear transformation on the fused composite feature tensor in the third hidden layer to enhance the regulatory effect of path information on depth prediction. is the path-aware fusion vector output by the third hidden layer, which contains information on path adjustment, nonlinear compression, and local feature integration; The final output crack area The depth prediction value of Used to characterize the longitudinal evolution degree of cracks in the image of the area; is the weight transpose vector of the output layer. The essence of the weight transpose vector is a dimension of A row vector of Represents the dimension of the path-aware fusion vector generated by the third hidden layer. The weight transposed vector is used to assign different linear contributions to each dimension of the previous layer output, that is, to compress high-dimensional features into a scalar prediction result through weighted summation. The value of comes from the back-propagation optimization process and is a trainable parameter. The model automatically adjusts each of its components by minimizing the prediction error.

[0181] is the bias vector of the output layer, and its dimension is , Consistent with the dimension of the path-aware fusion vector, it is used to introduce a fixed offset for each feature dimension after weighting to improve the nonlinear fitting ability of the model. Although the final linear mapping output is a scalar, the bias term retains its correspondence with the feature dimension in the form of a column vector. As a trainable parameter of the model, it is also iteratively optimized during the gradient descent process. The size of is, but is not limited to, initialized to zero or a small random value;

[0182] The state propagation representation tensor of the output of the evolutionary grid field layer After the structural channel decoupling operation, the input is respectively input into the crack morphology channel , Material Response Channel and path evolution channel , respectively generate channel feature tensors 、 、 ; The path-aware fusion vector is formed by fusion of feature superposition and path weight redistribution operation;

[0183] ;

[0184] ;

[0185] ;

[0186] in 、 、 are the dynamic attention coefficients in the process of modeling crack morphology, material response, and path evolution, respectively; 、 、 It can be generated through the learnable parameter matrix introduced in the forward propagation process. Specifically, the model dynamically adjusts the importance of each type of feature channel according to the prediction error through the backpropagation algorithm during the training phase. This is reflected in the local response weights of the three types of information (crack morphology, material response, and path evolution) at different image positions, thereby achieving differentiated modeling and fine feature fusion at the channel level. Represents the gradient difference based on the crack boundary and positional encoding Constructed edge feature coupling function; Represents material density distribution and the response nonlinear coefficient Material response mapping function; Represents a path-based direction tensor With feedback status label Constructed path evolution function; 、 、 They are channel activation and normalization functions, including functions such as ReLU and GELU, which are used to enhance the nonlinear expression ability of the channel and unify the feature scale to adapt to subsequent fusion calculations.

[0187] An image processing system for beam crack depth based on MLP optimization, including a labeling module, an evolution module, a guided expression module, a mutation module, and a prediction output module;

[0188] The annotation module is used to convert the original image data into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization, and generate a set of structurally annotated images with crack area labels through structural calibration and spatial mask construction;

[0189] The evolution module is used to construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation;

[0190] The guided expression module is used to map the image processing evolution table to the crack area according to the spatial position and construct a set of prediction probability distribution vectors. It also generates an expression vector with a spatial guidance mechanism through entropy density calculation and direction field construction, which serves as the input feature of the variant MLP model.

[0191] The mutation module is used to construct an image data processing structure based on a neural network: the hidden layer of the MLP model is divided into three layers, and a mutation MLP model is formed by introducing an evolutionary grid layer and a path control mechanism;

[0192] The prediction output module is used to input the expression vector with a spatial guidance mechanism into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of structural channel decoupling, and the third hidden layer of path regulation in sequence, and then perform state propagation, feature superposition, and path control weight regulation operations, and finally generate the depth prediction value of the corresponding crack area in the output layer.

[0193] It should be noted that this solution has evolved gradually in response to the problems existing in existing crack depth prediction technologies. Its core goal is to build a crack depth prediction model that integrates the image processing evolution mechanism with the neural network structure variation mechanism to achieve improved image data processing capabilities for crack areas. In this process, the overall design of the solution is carried out in a hierarchical manner according to five modules:

[0194] In S1, the solution starts with the original image acquired by the image acquisition device. It first suppresses noise through Gaussian filtering, and then uses grayscale mapping and contrast remapping techniques to enhance the saliency of crack boundaries. This stage aims to ensure that the image data has good background interference suppression capabilities and crack structure enhancement effects. Subsequently, through spatial cropping and size normalization, the image expression at the spatial distribution and scale levels is unified to avoid prediction bias caused by heterogeneous image sources. On this basis, the solution uses regional structure calibration operations to identify and delineate crack areas in blocks. At the same time, combined with spatial mask construction, a set of structurally annotated images with spatial reference and semantic directionality is generated, providing clear regional guidance information for subsequent state modeling.

[0195] Entering the S2 part, the scheme applies feature mapping operations to the set of structurally annotated images to construct an initial spatial state tensor. The tensor not only includes image grayscale and edge gradient information, but also embeds the mapping value of the structural label to reflect the fusion expression of local semantics and spatial layout. Based on this state tensor, a diffusion-driven propagation mechanism is further designed to simulate the process of crack features evolving layer by layer as the image space moves. The first evolutionary state map is generated to capture the initial propagation trend of the crack. The map repairs the path continuity through the directional gradient reconstruction operation and introduces the structural prior effect mapping in combination with the material property distribution function to construct the second evolutionary state map to reflect the potential guiding role of the material on the crack growth path. On this basis, the sequence stepping mechanism is used to iteratively simulate the multi-stage state expansion to obtain the state evolution sequence tensor, and finally through compression aggregation to form a deep spatial state expression vector, that is, the image processing evolution table, in preparation for further prediction.

[0196] The S3 part focuses on feature extraction and expression reconstruction before deep prediction based on the image processing evolution table. The semantic state embedding features in the evolution table are first mapped to the corresponding crack area according to the spatial position, and the continuous value prediction distribution fitting is performed on the state features in each area to form multiple regional prediction probability distribution vectors. After the local entropy density is calculated, these distribution vectors are used to construct the regional prediction entropy tensor, thereby identifying the uncertainty differences in the state in different areas of the image. The entropy tensor is further used to form an entropy change direction field through spatial gradient calculation to describe the flow direction of information entropy in the image. The direction field is combined with the evolution table to perform path vector fusion processing to extract highly indicative path guidance features. Finally, an expression vector with a spatial guidance mechanism is constructed through a step-by-step feedback superposition mechanism. This expression vector has adaptability to spatial heterogeneous structures and serves as the input of the neural network model.

[0197] In the S4 section, the proposal constructs a neural network framework based on the MLP model structure, and forms a mutated MLP model with directional adaptability through structural layer mutation. While maintaining the input and output layer structures, the model replaces the traditional first hidden layer with an evolutionary grid field layer to simulate the spatial propagation behavior of the structural state in the image, and establishes a state-driven logic based on spatial continuous expression. The second hidden layer adopts a structural channel decoupling method, which is divided into a crack morphology channel, a material response channel, and a path evolution channel. Independent mapping functions are constructed for structural expression, material medium, and path mechanism respectively. The three are integrated into a composite feature tensor through feature superposition, and then nonlinear activation and path control weight allocation operations are applied to achieve fusion expression. The third hidden layer introduces a path control weight regulation mechanism while retaining the fully connected structure to adjust the contribution of different feature paths to output prediction, further enhancing the model's responsiveness to path-guided structures.

[0198] Section S5 describes the complete inference process of the model. First, the spatially guided expression vector is input into the model input layer. After the initial projection mapping is completed, the input projection tensor is formed. This tensor is input into the evolution grid field layer, where the state propagation process is performed to form a state propagation representation tensor. The representation tensor then enters the decoupled three-channel structure for independent feature mapping. The three types of feature tensors output are integrated through feature superposition, and are processed by activation functions and path weights are adjusted to generate a path-aware fusion vector. This vector is input into the third hidden layer to perform full connection and path control operations, and finally a regression representation vector is generated. The output layer completes the linear mapping to generate the depth prediction value of the corresponding crack area, realizing end-to-end regression.

[0199] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An image processing method for beam crack depth based on MLP optimization, comprising acquiring raw image data through an image acquisition device, characterized in that: S1. The original image data is converted into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization. Then, a set of structurally annotated images with crack area labels is generated through structural calibration and spatial mask construction. S2. Construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation; S3. Map the image processing evolution table to the crack area according to the spatial position to construct a set of prediction probability distribution vectors. Then, through entropy density calculation and direction field construction, an expression vector with a spatial guidance mechanism is generated as the input feature of the variant MLP model. S4. Construct an image data processing structure based on a neural network: Divide the hidden layer of the MLP model into three layers, and form a variant MLP model by introducing an evolutionary grid layer and a path control mechanism; S5. The expression vector with the spatial guidance mechanism is sequentially input into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of the structural channel decoupling, and the third hidden layer of the path control. The state propagation, feature superposition, and path control weight control operations are performed in sequence, and finally the depth prediction value of the corresponding crack area is generated in the output layer.

2. The image processing method for beam crack depth based on MLP optimization according to claim 1, characterized in that: In S1, original image data is acquired through an image acquisition device, and an image data set including a pixel matrix and spatial position information is formed based on the original image data. The image data set is subjected to a noise suppression process by Gaussian filtering to form a background interference suppressed image set. The background interference suppressed image set is processed through grayscale mapping and contrast remapping to generate a crack boundary information enhanced image set. The crack boundary information enhanced image set is spatially cropped and size normalized to form an image set with uniform spatial distribution. The image set is then converted into a structure-annotated image set through regional structure calibration and spatial mask construction operations. The structure-annotated image set includes crack region block labels and spatial structure references.

3. The image processing method for beam crack depth based on MLP optimization according to claim 2, characterized in that: In S2, a set of structurally annotated images is obtained, and an initial spatial state tensor is generated from the set of structurally annotated images through a feature mapping operation. The initial spatial state tensor includes a local edge gradient, a grayscale value, and a structural label mapping value. The initial spatial state tensor is then state-advanced through a diffusion-driven propagation mechanism to generate a first evolutionary state map of the crack. The first evolutionary state map is reconstructed by performing path continuity reconstruction through a directional gradient reconstruction operation to form a crack direction state map. The crack direction state map is combined with a material property distribution function to perform structural prior effect mapping to generate a second evolutionary state map under the influence of material structure. The second evolutionary state map constructs a multi-step state evolution trajectory through a sequence stepping mechanism to form a state evolution sequence tensor; the state evolution sequence tensor is compressed and aggregated to construct a deep space state expression vector as an image processing evolution table.

4. The image processing method for beam crack depth based on MLP optimization according to claim 3 is characterized by: In S3, an image processing evolution table is obtained. After the image semantic state embedding features in the image processing evolution table are mapped to the corresponding crack areas according to the spatial positions, continuous value prediction distribution fitting is performed respectively to construct a set of prediction probability distribution vectors of the crack areas; the set of prediction probability distribution vectors is processed by local entropy density calculation to form a regional prediction entropy tensor, and the regional prediction entropy tensor is used to identify the degree of information uncertainty of each image area; The regional prediction entropy tensor is used to form an entropy change direction field through gradient calculation, and the entropy change direction field is used to describe the flow direction of information uncertainty in space; The entropy change direction field and the image processing evolution table are fused through path vector processing to construct a path guidance feature vector. The path guidance feature vector is subjected to a step-by-step feedback superposition operation to generate an expression vector with a spatial guidance mechanism. The expression vector is used as the input feature of the variant MLP model.

5. The image processing method for beam crack depth based on MLP optimization according to claim 4 is characterized in that: S4 also includes an MLP model and a variant MLP model; the MLP model includes an input layer, a hidden layer, and an output layer; the hidden layer of the MLP model is divided into a first hidden layer, a second hidden layer, and a third hidden layer; a variant MLP model is constructed based on the MLP model, the variant MLP model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; The input layer of the variant MLP model is used to receive the expression vector with a spatial guidance mechanism and complete the initial projection mapping; The first hidden layer of the mutated MLP model is replaced by an evolving lattice field layer, which performs state construction by simulating the spatial propagation process of structural states in the image. The second hidden layer of the variant MLP model is divided into a crack morphology channel, a material response channel, and a path evolution channel through structural channel decoupling, and performs independent feature mapping respectively; The third hidden layer of the variant MLP model retains the fully connected structure and introduces a path control weight adjustment operation; The output layer of the variant MLP model is a linear mapping structure, which is used to output the continuous prediction value of the crack depth; The variant MLP model is jointly constructed by a structural state propagation mechanism and a path information regulation mechanism, and is used to establish a mapping relationship from image features to crack depths.

6. The image processing method for beam crack depth based on MLP optimization according to claim 5, characterized in that: In S5, the expression vector with spatial guidance mechanism is input into the input layer of the variant MLP model and forms the input projection tensor after initial weight transformation; The input projection tensor is input to the first hidden layer, and the state propagation processing is performed through the evolved lattice mechanism of the evolved lattice layer, generating a state propagation representation tensor in the process of structural information diffusion and expression reconstruction; The state propagation representation tensor undergoes a decoupling transformation through the structural channel and enters the crack morphology channel, material response channel, and path evolution channel to perform independent feature mapping. The feature tensors of the three channels are output respectively. The feature tensors of the three channels are fused through feature superposition to construct a composite feature tensor. The second hidden layer performs activation function nonlinear transformation and path weight redistribution processing on the composite feature tensor to generate a path-aware fusion vector. The path-aware fusion vector is input into the third hidden layer, and the fully connected mapping and path control weight adjustment operations are performed to generate the final regression representation vector; The final regression representation vector is input to the output layer, where the depth prediction value of the corresponding crack area is generated through linear mapping.

7. The image processing method for beam crack depth based on MLP optimization according to claim 6, characterized in that: In S2, the structure annotation image collection Through the feature mapping function Mapped to the initial spatial state tensor , Including local edge gradient, gray value and structure label mapping value; using diffusion driven propagation mechanism to Execute the state advancement operation to simulate the crack evolution and expansion behavior and generate the first evolution state map ; Directed gradient reconstruction operation , perform path continuity recovery and generate crack direction state map ;Will Fusion Material Property Distribution Function and structural prior mapping fields , generating a second evolutionary state map under the influence of material structure ; Then through the sequence stepping mechanism Multi-step advancement to build state evolution sequence tensor ; will eventually Perform compression aggregation to generate deep spatial state expression vectors , as the image processing evolution table; Perform feature mapping and state tensor generation: ； ； ； Diffusion-driven first-order state evolution: ； Perform directional gradient reconstruction operation: ； Expression of structural prior mapping effect: ； Expression for multi-step state advancement: ； Compression aggregation into deep expression vector: ； The structure annotation image collection Used to provide spatial identification and label information of the crack area; feature mapping function Used to 、 and The joint encoding is the initial spatial state tensor; is the gradient tensor of the image; image grayscale normalization function Used to convert grayscale images The pixel values ​​in are linearly mapped to interval; is the crack area label mapping function, which is an indicator function used to map the crack area label at each pixel. The position of the crack area is marked as belonging to the crack area, and the output of the crack area label mapping function is Binary response plot of ; is the initial spatial state tensor; image Represents the pixel grayscale matrix of the original input image; Representing an image The lower limit of all pixel grayscale values ​​in ; Representing an image The upper limit of all pixel grayscale values ​​in ; is a Boolean-valued function; Representing an image Pixels in Whether it belongs to the crack area set ; The local diffusion coefficient tensor Indicates time Initial spatial state tensor function, Indicates time The initial spatial state tensor under , Used to characterize the information flow ability of each position in the image, The calculation form is: ,in is the diffusion inhibition factor, represents the square of the Euclidean two-norm, Express Perform directional gradient reconstruction operation, represents the spatial gradient operator, is an exponential function; is the global diffusion rate constant during the diffusion process; time For historical events, time is the current time; The termination diffusion time point of the first-level evolution process; Indicates time Approaching When , the final convergence value; is the first evolution state spectrum; the symbol · in the formula is the dot product; The directional gradient reconstruction operation In the formula, it is expressed as the directional gradient reconstruction function; where * is the convolution operation, is the component of the spatial gradient operator in the main direction; is the component of the spatial gradient operator in the vertical direction; is a directional convolution kernel, which is used to highlight the main direction structural features of the crack. The calculation form is: ； The direction is the local main direction angle; For direction Vertical convolution kernel; is the crack direction state map; in is the material property distribution function; the structural prior mapping field In the formula, it represents the structural prior mapping function; is the structural attribute modulation coefficient; is the second evolutionary state map; in The second evolutionary state map In time The evolutionary state of It is the state evolution promotion function, which is used to perform the update operation of each state step. The calculation form is: ； in is the Laplace operator, the hyperbolic tangent function In the formula, represents the nonlinear saturation transformation; is the diffusion control coefficient; is the structural attribute modulation coefficient; is the time step of state evolution; is the state evolution sequence tensor, which records the state evolution sequence tensor from to All state evolution graphs of ; is the time series length of the second evolution stage; in is the tensor compression aggregation function; is the activation function; is the global average pooling function; For time The weight coefficient of the step state in the aggregation; Evolution table vector for image processing.

8. The image processing method for beam crack depth based on MLP optimization according to claim 7, characterized in that: In S3, it also includes: ； ； ； ； ； ； in The crack area Medium pixel The predicted probability distribution value of ; For the region Normalization factor on ; Pixel points in the image processing evolution table Status characteristics; Pixel The predicted target value mapping; The center pixel of the crack With pixels The Euclidean distance between is the exponential scaling factor of the forecast difference; is the distance adjustment coefficient; in Indicates the The set of all pixels in the crack area; Pixel The image gradient vector at ; Represents pixel points and The square of the second norm of the image gradient difference; Pixel The information entropy direction field vector; Pixel The divergence of the information entropy direction field at ; is the exponential control factor, defined as: ; Indicates the angle; is the crack area label mapping function, if the pixel Belong to the crack area, then ,otherwise ; is the crack density function, defined as: , Represents pixel points Neighborhood window of Represents pixel points A pixel in the neighborhood of ; Represents pixel points Crack area label mapping function; in The crack area The predicted entropy density value of ; It is the minimum protection constant in entropy calculation, which is used to prevent the logarithmic term from going to infinity; is the direction vector of the regional entropy gradient; is the symbol of partial derivative; is the horizontal coordinate axis of the image in the two-dimensional plane, is the vertical coordinate axis of the image in the two-dimensional plane; in Crack area in the image processing evolution table The original path feature vector of ; is the path history response operator; Represents the element-wise product between tensors; is the fusion operator; The crack area The guided path vector of in The crack area The spatial guide expression vector of Finally, it serves as the input of the variant MLP model; For the feedback superposition Step adjustment weight of step; is the preset feedback step constant; is the activation function; is the step response function; is the step index of the current jump; Indicates the total number of steps of skip feedback superposition.

9. The image processing method for beam crack depth based on MLP optimization according to claim 8, characterized in that: Construct the mutation MLP model in S4 and S5: Input layer: ； First hidden layer: ； Second hidden layer: ； in: ； Third hidden layer: ； Output layer: ； in is the input mapping function; is the input feature vector after projection; is the evolution lattice function; is the weight mapping matrix of the input layer; is the bias vector of the input layer; The crack area The spatial neighborhood set of ; From the crack area To the crack area Propagation factor of crack area Indicates the current crack area In, with Indexes of adjacent pixels or regions that have spatial or path connectivity; is the grid field kernel function; The state propagation representation tensor output by the evolutionary grid field layer; They are the feature extraction mapping functions for the crack morphology channel, material response channel, and path evolution channel respectively; Independent feature tensors output for three types of channels; is the composite feature tensor after channel fusion; is the path regulation mapping matrix of the third hidden layer; represents the information guided weight function; Indicates crack area superior About Pixels gradient; is the activation control function of the path regulation layer; is the weight mapping matrix of the path control layer; is the path-aware fusion vector output by the third hidden layer; The final output crack area The depth prediction value of is the weight transpose vector of the output layer; is the bias vector of the output layer.

10. An image processing system for beam crack depth based on MLP optimization, comprising a labeling module, an evolution module, a guided expression module, a mutation module, and a prediction output module, characterized in that: The annotation module is used to convert the original image data into a set of images with uniform spatial distribution through noise suppression, contrast remapping and size normalization, and generate a set of structurally annotated images with crack area labels through structural calibration and spatial mask construction; The evolution module is used to construct a state evolution sequence tensor from the structure-annotated image set through feature mapping, state advancement, and three-level evolution processing, and generate an image processing evolution table through compression and aggregation; The guided expression module is used to map the image processing evolution table to the crack area according to the spatial position and construct a set of prediction probability distribution vectors. It also generates an expression vector with a spatial guidance mechanism through entropy density calculation and direction field construction, which serves as the input feature of the variant MLP model. The mutation module is used to construct an image data processing structure based on a neural network: the hidden layer of the MLP model is divided into three layers, and a mutation MLP model is formed by introducing an evolutionary grid layer and a path control mechanism; The prediction output module is used to input the expression vector with a spatial guidance mechanism into the input layer of the variant MLP model, the first hidden layer of the evolutionary grid field, the second hidden layer of structural channel decoupling, and the third hidden layer of path regulation in sequence, and then perform state propagation, feature superposition, and path control weight regulation operations, and finally generate the depth prediction value of the corresponding crack area in the output layer.

Images