GIL device defect recognition method based on x-ray image enhancement and semantic segmentation

Abstract

This invention relates to the field of equipment defect identification, and more particularly to a GIL (Gas Injection Line) equipment defect identification method based on X-ray image enhancement and semantic segmentation. First, the acquired multi-energy spectral response values ​​are normalized to obtain normalized energy spectral response values. Based on these normalized values, the acquired grayscale image data is enhanced to obtain enhanced image data. Then, based on the enhanced image data, structural correction and feature extraction are performed to obtain feature vectors. These feature vectors are standardized and mapped, and nonlinear compression and weighting are introduced to obtain weighted feature vectors. Feature classification is then performed, and class probabilities are calculated. Based on these probabilities, defect feature vectors are obtained. Finally, based on these feature vectors, cosine similarity is calculated, and the class corresponding to the maximum cosine similarity is selected to determine the equipment defect. This method solves the technical problems of difficult effective separation of defect features, inconsistent structural information, and unstable identification results.

Description

Technical Field

[0001] This invention relates to the field of equipment defect identification, and more particularly to a GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation. Background Technology

[0002] As power systems evolve towards higher voltage, larger capacity, and higher reliability, gas-insulated metal-enclosed transmission line (GIL) equipment is widely used in critical scenarios such as large substations and hydropower stations due to its advantages of small footprint, high operational stability, and strong environmental adaptability. However, this type of equipment operates in a closed, high-voltage environment for extended periods, and its internal structure is complex, consisting of conductors, insulators, a shell, and various connecting components. Assembly defects, material defects, or foreign object defects can easily trigger partial discharges or even lead to equipment failure. Therefore, accurate detection of internal defects is of great significance. In existing technologies, X-ray inspection, as a typical non-destructive testing method, can obtain information about the internal structure of the equipment without disassembling it and has been gradually applied to defect detection in GIL equipment. By allowing X-rays to penetrate the internal structure of the equipment and form a grayscale image on the detector, the differences in the absorption of X-rays by different materials can be reflected, thus enabling preliminary identification of internal defects. However, since GIL equipment is usually composed of multiple materials stacked together, the attenuation characteristics of different materials differ, leading to grayscale overlap during imaging. This results in low contrast between the defect area and the background structure, especially making it difficult to clearly present defects such as micro-cracks and bubbles in the original image. Therefore, there is an urgent need to propose a defect identification method for GIL equipment based on X-ray image enhancement and semantic segmentation to solve the above problems. Summary of the Invention

[0003] This invention provides a defect identification method for GIL equipment based on X-ray image enhancement and semantic segmentation, in order to solve the technical problems of difficulty in effectively separating defect features, inconsistent structural information, and unstable identification results caused by the superposition of multiple material structures, imaging noise interference, and differences in the operating conditions of different equipment during the X-ray inspection of GIL equipment.

[0004] The present invention provides a GIL (Gas Injection Line) equipment defect identification method based on X-ray image enhancement and semantic segmentation, comprising the following steps:

[0005] S1. Obtain the original data containing grayscale image data and the corresponding pixel position multi-spectral response values. Normalize the multi-spectral response values ​​to obtain normalized energy spectrum response values. Based on the normalized energy spectrum response values, enhance the grayscale image data by constructing a multi-spectral fusion enhancement response factor and introducing a spatial gradient suppression term to obtain enhanced image data.

[0006] S2. Based on the enhanced image data, perform structural correction processing to obtain a structurally corrected image; based on the structurally corrected image, perform feature extraction to obtain a feature vector; after standardizing the feature vector, map it to the target domain, and further introduce nonlinear compression to obtain a compressed feature vector; based on the compressed feature vector, perform weighted processing to obtain a weighted feature vector; based on the weighted feature vector, perform feature classification to obtain a feature classification set; based on the feature classification set, calculate the category probability; based on the category probability, perform defect region division and feature extraction to obtain a defect feature vector; based on the defect feature vector and the obtained defect sample feature vector, perform cosine similarity calculation, select the category corresponding to the maximum cosine similarity as the output, and determine the equipment defect.

[0007] Preferably, S1 specifically includes:

[0008] The normalized spectral response values ​​include those of the low-energy, mid-energy, and high-energy channels. The low-energy enhancement index is calculated based on the standard deviation of the grayscale image data plus the reciprocal of 1. The material sensitivity ratio is calculated using a power function, combining the normalized spectral response values ​​of the low-energy and mid-energy channels with the low-energy enhancement index. Finally, the multi-spectral fusion enhancement response factor is calculated based on the material sensitivity ratio and the normalized spectral response value of the high-energy channel.

[0009] Preferably, S1 specifically includes:

[0010] Based on grayscale image data, local gradient calculation is performed on the grayscale values ​​of the grayscale image data to obtain the gradient of the pixel; based on the gradient of the pixel, the gradient magnitude is calculated; based on the gradient magnitude, a spatial gradient suppression term is constructed using a sine function.

[0011] Preferably, S1 specifically includes:

[0012] The effective transmission intensity after multi-spectral modulation is calculated by multiplying the grayscale image data with the multi-spectral fusion enhancement response factor. The spatial gradient suppression term is divided by 1 and the multi-spectral fusion enhancement response factor is added to calculate the relative perturbation level. Based on the relative perturbation level, a suppression coefficient is introduced to calculate the structural credibility modulation factor. The enhanced image data is obtained by multiplying the effective transmission intensity after multi-spectral modulation with the structural credibility modulation factor.

[0013] Preferably, S2 specifically includes:

[0014] The relative error is calculated based on the absolute value of the difference between the pixel grayscale values ​​of the enhanced image data and the set structural reference image. The grayscale deviation is calculated by dividing the relative error by 1 and adding the pixel grayscale value of the structural reference image. The local gradients of the enhanced image and the structural reference image are calculated based on the pixel grayscale values ​​of the enhanced image data and the structural reference image, respectively. The direction angle is calculated using the inverse cosine function based on the local gradients of the enhanced image and the structural reference image. The structural weight is calculated based on the direction angle and the grayscale deviation. The structural correction image is obtained based on the structural weight and the enhanced image data.

[0015] Preferably, S2 specifically includes:

[0016] Based on the feature vector, the feature vector is standardized by the mean and standard deviation of the feature vector, and the standardized feature vector is calculated. Based on the obtained standard equipment features, after normalization, the standard deviation and mean are calculated respectively to obtain the feature standard deviation and feature mean.

[0017] Preferably, S2 specifically includes:

[0018] Based on the feature standard deviation, feature mean, feature vector, and standardized feature vector, a mapping is performed to obtain the mapped feature vector; based on the mapped feature vector, nonlinear compression is introduced to obtain the compressed feature vector.

[0019] Preferably, S2 specifically includes:

[0020] Based on the compressed eigenvectors, the sigmoid function term is calculated; based on the compressed eigenvectors, a structure enhancement adjustment coefficient is introduced to calculate the structure enhancement factor; the structure enhancement factor is multiplied by the sigmoid function term to calculate the dynamic weights; the dynamic weights are multiplied by the compressed eigenvectors to obtain the weighted eigenvectors.

[0021] Preferably, S2 specifically includes:

[0022] Based on the category probability, the category with the highest probability value is selected as the initial label of the pixel, thereby generating an initial classification label map. Based on the set probability threshold, the initial classification label map is subjected to confidence constraint processing, and the initial labels of pixels with category probabilities greater than or equal to the probability threshold are retained. After further morphological processing using digital image morphology methods, the set of pixels with the same category and spatial continuity is divided into independent regions.

[0023] Preferably, S2 specifically includes:

[0024] Based on the set area threshold, excessively small regions in the independent regions are removed to obtain defect segmentation regions; feature extraction is performed on the defect segmentation regions to obtain defect feature vectors.

[0025] The beneficial effects of the technical solution of the present invention are:

[0026] 1. By introducing multi-energy spectral response values ​​and X-ray attenuation laws, the original grayscale image is expanded from a single intensity expression to a composite expression of "grayscale-energy spectrum-material properties." Normalization eliminates amplitude deviations caused by different acquisition conditions, allowing subsequent calculations to be performed on a uniform scale, avoiding the recognition instability caused by exposure differences in traditional methods. By constructing a multi-energy spectral fusion enhancement response factor, the high sensitivity of the low-energy spectrum, the stability of the medium-energy spectrum, and the penetrability of the high-energy spectrum are synergistically utilized, significantly amplifying material differences in the enhanced image data, thereby effectively improving the distinguishability of minute defects such as cracks and bubbles. Simultaneously, by introducing a spatial gradient suppression term, high-frequency noise features are distinguished from the true structural boundaries, achieving a dual effect of "defect enhancement and noise suppression" during the enhancement process. This ensures that the enhanced image possesses both high contrast and maintains structural continuity, providing high-quality input data for subsequent processing.

[0027] 2. In the structural correction stage, by jointly modeling grayscale deviation and orientation angle, a pixel-level correspondence is established between the enhanced image and the features of the standard device. Furthermore, boundary directions are constrained at the spatial structure level, ensuring that each pixel possesses both "grayscale consistency" and "orientation consistency" attributes in the output structurally corrected image. Through structural weight mapping, regions conforming to structural features are enhanced, while abnormal regions, such as defect locations, are naturally highlighted within the overall structure, significantly reducing the interference of complex backgrounds on recognition. This process allows subsequent feature extraction to no longer rely on the unstable structure of the original image, but rather on the corrected, standardized structural space, improving the stability and consistency of feature representation. Attached Figure Description

[0028] Figure 1 This is a flowchart of the GIL device defect identification method based on X-ray image enhancement and semantic segmentation as described in this invention. Detailed Implementation

[0029] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] The following description, in conjunction with the accompanying drawings, details the specific scheme of the GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation provided by this invention.

[0032] See attached document Figure 1 The diagram illustrates a flowchart of a GIL device defect identification method based on X-ray image enhancement and semantic segmentation according to an embodiment of the present invention. The method includes the following steps:

[0033] S1. Obtain the original data containing grayscale image data and the multi-energy spectrum response values ​​of the corresponding pixel positions. Normalize the multi-energy spectrum response values ​​to obtain normalized energy spectrum response values. Based on the normalized energy spectrum response values, enhance the grayscale image data by constructing a multi-energy spectrum fusion enhancement response factor and introducing a spatial gradient suppression term to obtain enhanced image data.

[0034] The internal structure of the GIL device was photographed using an existing X-ray multispectral imaging device to acquire raw data containing two-dimensional grayscale image data and multi-energy spectral response values ​​of corresponding pixel locations. This raw data was then used as the basic input data set and denoted as a triplet. : , in, Represents the original two-dimensional grayscale image in pixels The grayscale value at that location, i.e., grayscale image data; , , These represent the response values ​​of the low-energy, medium-energy, and high-energy channels, respectively, i.e., the transmission intensities of different energy channels. The low-energy channel has an energy range of approximately 20-80 keV and is used to detect minute defects such as cracks and bubbles. The medium-energy channel has an energy range of approximately 80-160 keV and is used for structure identification and localization. The high-energy channel has an energy range of approximately 160-300 keV and is used for overall structure modeling. The transmission intensities of different energy channels are used to describe the X-ray penetration results under different accelerating voltage conditions. Since X-ray attenuation in different materials follows the X-ray attenuation law, the response values ​​of different energy channels satisfy: ,in, Indicates the first Under certain energy spectrum conditions, after the X-rays penetrate the object being measured, they reach the detector pixels. The transmitted intensity received at point , i.e., the first One energy spectrum response value; It is the incident X-ray intensity, that is, the initial intensity of the rays before they penetrate an object, and is the calibration value of an X-ray multispectral imaging device; It is the first Under the condition of individual energy spectrum, at pixel The linear attenuation coefficient on the corresponding path is determined according to the material type. For example, the value is 5 to 50 for copper or aluminum, and 0.5 to 5 for epoxy resin (insulation). This is the equivalent path length of a ray propagating through an object, obtained using existing ray tracing algorithms such as Ray-tracing. The number of rays is one per pixel, the maximum recursion depth is set to 1, and the ray step size is in the range of 1mm to 5mm. Furthermore, the spectral response values ​​are normalized using the maximum value method to obtain the normalized spectral response values. This eliminates amplitude differences caused by different acquisition conditions, enabling subsequent calculations to be based on a unified scale.

[0035] Furthermore, a multi-spectral fusion-enhanced response factor is constructed, based on the following: low-energy spectra are sensitive to material density and require amplification; mid-energy spectra serve as a stable reference for suppression; and high-energy spectra have strong penetration but low contrast, necessitating compression of their dynamic range. Therefore, the specific calculation formula for the multi-spectral fusion-enhanced response factor is as follows: , in, It is a multi-spectral fusion enhanced response factor; It is the low-energy spectrum enhancement index, used to control the degree of low-energy spectrum enhancement. It is determined based on contrast, that is, the reciprocal of the standard deviation of the gray values ​​of the original grayscale image data plus 1 is used as the low-energy spectrum enhancement index. This is a stability constant used to prevent the denominator from being zero, such as... ; This represents the normalized spectral response value of the low-energy channel; This represents the normalized spectral response value of the medium energy channel; This represents the normalized spectral response value of the high-energy spectral channel; This represents the material sensitivity ratio, used to highlight material differences while suppressing the impact on the overall structure. This indicates an "enhancement-compression" process applied to the penetration information of the high-energy spectrum, allowing deep structural information to participate in the fusion without compromising numerical stability. The multi-energy spectral fusion enhancement response factor calculation formula uses a power function to boost weak signals, a square root to achieve smooth suppression, and a logarithmic function to compress the dynamic range of the high-energy spectrum. This combination originates from the inverse reconstruction of the exponential decay model.

[0036] Next, to avoid amplifying noise during the enhancement process, a spatial gradient suppression term is introduced. First, the pixel grayscale values ​​of the original two-dimensional grayscale image are processed using the existing center difference method. Perform local gradient calculation to determine the pixel's position. and gradient in direction , Thus, the gradient magnitude is obtained. This reflects the degree of drastic local changes. Since noise typically exhibits high-frequency variations, a spatial gradient suppression term is constructed. Its construction logic is as follows: when the gradient is small, the fraction is close to linear, preserving details; when the gradient is large, it tends to saturate, thus avoiding excessive suppression of true edges.

[0037] The final enhancement result is: , in, It is an enhanced X-ray image at the pixel level. The grayscale value at that location is used to represent the enhanced image data; This is the suppression coefficient, used to control the proportion of noise suppression in the overall enhancement. It is determined using the five-fold cross-validation method, and the reference value range is [value missing]. ; It is the effective transmission intensity after multi-spectral modulation; It is the structural credibility modulation factor; Indicates the relative degree of perturbation. The enhanced image data... As the sole input for the next stage.

[0038] S2. Based on the enhanced image data, perform structural correction processing to obtain a structurally corrected image; based on the structurally corrected image, perform feature extraction to obtain a feature vector; after standardizing the feature vector, map it to the target domain, and further introduce nonlinear compression to obtain a compressed feature vector; based on the compressed feature vector, perform weighted processing to obtain a weighted feature vector; based on the weighted feature vector, perform feature classification to obtain a feature classification set; based on the feature classification set, calculate the category probability; based on the category probability, perform defect region division and feature extraction to obtain a defect feature vector; based on the defect feature vector and the obtained defect sample feature vector, perform cosine similarity calculation, select the category corresponding to the maximum cosine similarity as the output, and determine the equipment defect.

[0039] The structural reference image is pre-set by technicians based on the physical model of X-ray attenuation (Beer-Lambert's law) and CAD geometric modeling. Further, the structural correction stage begins, with the enhanced image data as input. and pixel grayscale values ​​of the structural reference image First, based on relative error Calculate grayscale deviation , Representing the structural reference image in pixels The grayscale value at that location.

[0040] Further, the directional angle of the pixels is calculated based on the pixel grayscale values ​​of the enhanced image. and the pixel grayscale values ​​of the structural reference image Local gradient calculations were performed to obtain the local gradients of the enhanced image and the structural reference image, respectively. , Then in pixels Angle of direction at the location for: , in, Represents pixels The directional angle at a point is used to describe the degree of consistency of the boundary directions; This is a stability constant used to prevent the denominator from being zero, such as... ; Indicates the enhanced image Gradient in direction; Indicates the enhanced image Gradient in direction; Representing structural reference images Gradient in direction; Representing structural reference images Gradient in the direction.

[0041] Combining grayscale deviation and orientation angle, a pixel-based structure is constructed. Structural weights at the location : , in, Indicated in pixels Structural weights at the location; ; The cosine term is used to describe directional consistency. Indicated in pixels Gray-scale deviation at the location; This is a fractional term used to describe grayscale deviation suppression.

[0042] Output structure correction image: ,in, It is the enhancement adjustment coefficient, determined using the five-fold cross-validation empirical method, with a reference range of values. .

[0043] Furthermore, regarding Feature extraction is performed using existing convolutional neural networks to obtain feature vectors. Specifically, a convolutional neural network consists of an input layer, two convolutional layers, two pooling layers, and a feature integration layer. It also includes a fully connected layer for feature classification. The input layer receives the structure-corrected image; the first convolutional layer has 32 pixels of size... The convolutional kernel has a stride of 1, uses the same padding method, and undergoes non-linear activation via ReLU to extract edge and basic texture features; subsequently, it is processed through a window... The first layer uses a max-pooling layer with a stride of 2 for downsampling to reduce spatial resolution and enhance robustness; the second convolutional layer has 64 steps. The convolutional kernel further extracts higher-level structural and texture information based on the features from the previous layer, and also uses the ReLU activation function; then, a second max-pooling layer is used for feature compression again; on this basis, a layer of... Convolution, used as a feature integration layer, has 128 kernels and performs linear combination and reconstruction while preserving the spatial structure, ultimately outputting a feature vector. During training, the convolutional neural network uses a learning rate of 0.001, a batch size of 16, and 50 training rounds. The cross-entropy loss function is used, and the Adam optimization algorithm is employed as the optimizer. This ensures that the feature extraction process has stable convergence and good generalization ability. The feature vector serves as the direct input to the subsequent feature normalization and semantic segmentation modules, forming a continuous data processing chain.

[0044] Furthermore, to address the issue of cross-device differences, it is necessary to refine the feature vectors. Standardize the vectors. First, calculate the mean of the eigenvectors. and standard deviation Then, standardization is performed to obtain the standardized feature vector, the specific expression of which is: , in, This represents the standardized feature vector; This represents the mean of the eigenvectors; The standard deviation of the eigenvectors; This is a stability constant used to prevent the denominator from being zero, such as... .

[0045] Then map to the target domain, the specific expression is: , in, It is the mapped feature vector; It is the characteristic fluctuation range under standard equipment, i.e. the characteristic standard deviation, which is used to control the characteristic contrast scale. It is obtained by retrieving standard equipment features from the existing standard equipment database and calculating the standard deviation after normalization such as the maximum value method. It is the mean value of features under standard equipment, which is obtained by retrieving the features of standard equipment from the existing standard equipment database and then calculating the mean value after normalization such as the maximum value method. It is an adaptive enhancement modulation term. "1" ensures that the basic linear mapping is not weakened, while the subsequent fractional terms enhance weak features and keep strong features stable, thereby achieving selective enhancement of defective regions. It is a cross-device alignment and nonlinear enhancement coupling term that performs intensity-related amplification of defect features under the condition of a unified reference device.

[0046] To prevent the influence of extreme values, nonlinear compression is introduced to obtain the compressed feature vector. : ,

[0047] The above formula compresses large values ​​using logarithms and enhances small values ​​using square roots, thus ensuring a balanced distribution.

[0048] Furthermore, the compressed feature vector Perform weighted processing. First, calculate the weighted average at the pixel level. Dynamic weights at the location : , in, Represents pixels Dynamic weights at each location; This is the structure enhancement adjustment coefficient, i.e., the attention gain factor, employing a lightweight channel attention mechanism. For example, if the SE module is defined, the number of heads is 16, the dimension is 256, and the reference value range is... ; This represents the sigmoid function term, used to provide probability constraints; It is a structure enhancement factor used to strengthen weak feature regions such as edges and cracks. Further, it is obtained at the pixel level. Weighted eigenvectors at location .

[0049] Furthermore, the weighted feature vector is processed through the fully connected layers in the aforementioned convolutional neural network. The process is performed to classify the features, resulting in a feature classification set. ,in, Indicates the first Weighted feature vectors of each category, This represents the index of the category. Further, the category probability is calculated using the Softmax function. ,in, Represents pixels The first The probability of each category.

[0050] Further, defect region segmentation is performed. Specifically, the probability of each category at each pixel location is compared, and the category with the highest probability value is selected as the initial label for that pixel, thus generating an initial classification label map. Based on this, the initial classification label map is subjected to confidence constraint processing. Pixels with category probabilities lower than the probability threshold preset by technicians based on the quantile threshold method are uniformly classified as background categories to reduce the interference of uncertain regions on the results. Pixels with category probabilities greater than or equal to the probability threshold retain their initial labels. Subsequently, existing mature digital image morphology methods are used for morphological processing, including existing erosion operations, such as 3×3 square structuring elements, to remove isolated noise points. After completing the above morphological processing, existing two-pass scanning algorithms or connected component algorithms based on disjoint-set data structures are used to divide the set of pixels with the same initial category label and spatial contiguousness into independent regions. Finally, based on the region area threshold, such as the minimum number of pixels, excessively small regions are eliminated, thus obtaining the final defect segmentation region.

[0051] Furthermore, the aforementioned convolutional neural network is used to extract features from the defect segmentation region to obtain the defect feature vector. The cosine similarity is calculated between the feature vector of the defect sample obtained from the existing standard equipment database and the feature vector of the defect sample. The category corresponding to the maximum cosine similarity is then selected as the output.

[0052] In summary, a defect identification method for GIL equipment based on X-ray image enhancement and semantic segmentation has been completed.

[0053] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0054] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying defects in GIL equipment based on X-ray image enhancement and semantic segmentation, characterized in that, Includes the following steps: S1. Obtain the raw data containing grayscale image data and the corresponding pixel position multi-energy spectrum response value, and normalize the multi-energy spectrum response value to obtain the normalized energy spectrum response value. Based on the normalized energy spectrum response value, the grayscale image data is enhanced by constructing a multi-energy spectrum fusion enhancement response factor and introducing a spatial gradient suppression term to obtain the enhanced image data. S2. Based on the enhanced image data, perform structural correction processing to obtain a structurally corrected image; based on the structurally corrected image, perform feature extraction to obtain a feature vector; After standardizing the feature vectors, they are mapped to the target domain, and then nonlinear compression is introduced to obtain compressed feature vectors. Based on the compressed feature vectors, weighting is performed to obtain weighted feature vectors. Based on the weighted feature vectors, feature classification is performed to obtain a feature classification set. Based on the feature classification set, the category probability is calculated; based on the category probability, the defect region is divided and features are extracted to obtain the defect feature vector; Based on the defect feature vector and the obtained defect sample feature vector, cosine similarity is calculated, and the category corresponding to the maximum cosine similarity is selected as the output to determine the equipment defect.

2. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 1, characterized in that, S1 specifically includes: The normalized spectral response values ​​include those of the low-energy, mid-energy, and high-energy channels. The low-energy enhancement index is calculated based on the standard deviation of the grayscale image data plus the reciprocal of 1. The material sensitivity ratio is calculated using a power function, combining the normalized spectral response values ​​of the low-energy and mid-energy channels with the low-energy enhancement index. Finally, the multi-spectral fusion enhancement response factor is calculated based on the material sensitivity ratio and the normalized spectral response value of the high-energy channel.

3. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 2, characterized in that, S1 specifically includes: Based on grayscale image data, local gradient calculation is performed on the grayscale values ​​of the grayscale image data to obtain the gradient of the pixel; based on the gradient of the pixel, the gradient magnitude is calculated; based on the gradient magnitude, a spatial gradient suppression term is constructed using a sine function.

4. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 3, characterized in that, S1 specifically includes: The effective transmission intensity after multi-spectral modulation is calculated by multiplying the grayscale image data with the multi-spectral fusion enhancement response factor. The spatial gradient suppression term is divided by 1 and the multi-spectral fusion enhancement response factor is added to calculate the relative perturbation level. Based on the relative perturbation level, a suppression coefficient is introduced to calculate the structural credibility modulation factor. The enhanced image data is obtained by multiplying the effective transmission intensity after multi-spectral modulation with the structural credibility modulation factor.

5. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 1, characterized in that, S2 specifically includes: The relative error is calculated based on the absolute value of the difference between the pixel grayscale values ​​of the enhanced image data and the set structural reference image. The grayscale deviation is calculated by dividing the relative error by 1 and adding the pixel grayscale value of the structural reference image. The local gradients of the enhanced image and the structural reference image are calculated based on the pixel grayscale values ​​of the enhanced image data and the structural reference image, respectively. The direction angle is calculated using the inverse cosine function based on the local gradients of the enhanced image and the structural reference image. The structural weight is calculated based on the direction angle and the grayscale deviation. The structural correction image is obtained based on the structural weight and the enhanced image data.

6. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 1, characterized in that, S2 specifically includes: Based on the feature vector, the feature vector is standardized by the mean and standard deviation of the feature vector, and the standardized feature vector is calculated. Based on the obtained standard equipment features, after normalization, the standard deviation and mean are calculated respectively to obtain the feature standard deviation and feature mean.

7. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 6, characterized in that, S2 specifically includes: Based on the feature standard deviation, feature mean, feature vector, and standardized feature vector, a mapping is performed to obtain the mapped feature vector; based on the mapped feature vector, nonlinear compression is introduced to obtain the compressed feature vector.

8. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 7, characterized in that, S2 specifically includes: Based on the compressed eigenvectors, the sigmoid function term is calculated; based on the compressed eigenvectors, a structure enhancement adjustment coefficient is introduced to calculate the structure enhancement factor; the structure enhancement factor is multiplied by the sigmoid function term to calculate the dynamic weights; the dynamic weights are multiplied by the compressed eigenvectors to obtain the weighted eigenvectors.

9. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 1, characterized in that, S2 specifically includes: Based on the category probability, the category with the highest probability value is selected as the initial label of the pixel, thereby generating an initial classification label map. Based on the set probability threshold, the initial classification label map is subjected to confidence constraint processing, and the initial labels of pixels with category probabilities greater than or equal to the probability threshold are retained. After further morphological processing using digital image morphology methods, the set of pixels with the same category and spatial continuity is divided into independent regions.

10. The GIL equipment defect identification method based on X-ray image enhancement and semantic segmentation according to claim 9, characterized in that, S2 specifically includes: Based on the set area threshold, excessively small regions in the independent regions are removed to obtain defect segmentation regions; feature extraction is performed on the defect segmentation regions to obtain defect feature vectors.

Images