Filter mesh hole defect intelligent identification method and system based on deep convolutional network

Abstract

The application provides a filter screen micro-hole defect intelligent identification method and system based on a deep convolution network, relates to the technical field of artificial intelligence, and comprises the following steps: different aperture micro-holes are processed through a scale self-adaptive pooling technology, cross-layer fusion of shallow geometric features and deep semantic features is combined, and the identification capability of boundary fuzzy defects is improved. The micro-hole spatial adjacency relationship analysis is introduced, the neighborhood information is used to assist in determining the micro-holes with confidence conflict, and the fusion weight parameters are iteratively optimized through correction feedback, so that high-precision and high-adaptability defect detection is realized.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to a method and system for intelligent identification of micropore defects in filter screens based on deep convolutional networks. Background Technology

[0002] As a critical power transmission system for aircraft, the hydraulic system's filter screen is a vital safety component ensuring its normal operation. It is responsible for intercepting tiny particles and impurities in the hydraulic oil, preventing them from entering precision hydraulic components and causing wear or blockage. Current inspection methods primarily rely on manual visual inspection or basic machine vision technology, which have significant limitations. Manual inspection is not only inefficient but also susceptible to fatigue, making it difficult to guarantee consistency. Existing machine vision methods perform poorly when dealing with the unique metallic mesh structure, complex light reflection, and densely packed micropores of hydraulic filters. In particular, after high-pressure pulse impacts or prolonged service, the micropore boundaries of hydraulic filters may deform or become blurred, significantly reducing the accuracy of traditional algorithms. Aviation safety standards require that hydraulic filter inspections cannot miss any potential dangerous defects, but current technologies struggle to control the false positive rate while maintaining a high detection rate. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent identification of micropore defects in filter screens based on deep convolutional networks, which can solve the problems in existing technologies.

[0004] A first aspect of this invention provides a method for intelligent identification of micropore defects in filter screens based on deep convolutional networks, comprising: A surface image of the metal filter screen of the aircraft hydraulic system to be inspected is acquired, and the surface image is subjected to illumination normalization and noise suppression processing to obtain a standardized image; Micropore contours are extracted from the standardized image based on morphological operators and connected component analysis rules, resulting in a set of single micropore candidate regions containing spatial location information and boundary morphology information. Multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. Furthermore, shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fused features are input into the classification decision module, which determines the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generates a confidence score for each defect type to obtain the defect identification result. Based on the spatial location information of each micropore candidate region in the defect identification results, a micropore spatial adjacency relationship graph is constructed. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the type correction is performed on the micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect type is fed back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the ability of the fusion feature to distinguish defects with blurred boundaries is improved, and a defect detection report is obtained.

[0005] In one optional implementation, the steps of acquiring a surface image of the metal filter to be inspected, and performing illumination normalization and noise suppression processing on the surface image to obtain a standardized image include: The illumination distribution pattern in the surface image is extracted based on the statistical characteristics of the gray-level histogram. The gray-level deviation between each pixel and the target illumination reference is calculated according to the illumination distribution pattern. The gray-level values ​​of each pixel in the surface image are adaptively compensated and corrected according to the gray-level deviation to eliminate regional brightness differences and obtain an illumination-normalized image. The illumination-normalized image is decomposed in the frequency domain to separate high-frequency noise components and low-frequency structural components. The high-frequency noise components are selectively suppressed based on edge-preserving constraints, while the micropore boundary information in the low-frequency structural components is preserved. The suppressed high-frequency components and the low-frequency structural components are then reconstructed and fused to obtain a normalized image.

[0006] In one optional implementation, the step of extracting micropore contours from the standardized image based on morphological operators and connected component analysis rules to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information includes: The standardized image is binarized to distinguish between the micropore region and the background region. Multi-scale morphological opening and closing operations are used to fill holes and smooth edges in the binarized image to remove pseudo-connected regions caused by metal surface texture and noise residue, thus obtaining a preliminary segmented image. The preliminary segmented image is labeled with connected components, and the geometric shape parameters and topological features of each connected component are extracted. Based on the geometric shape parameters and topological features, the adhesion micropore regions are identified and separated. The boundaries of the identified adhesion micropore regions are reconstructed based on the principle of indentation point detection and watershed segmentation to determine the independent contour boundaries of each individual micropore. Traverse the independent contour boundaries, extract the center coordinates of each independent contour boundary as spatial location information, and extract the perimeter, area, roundness and major-minor axis ratio of each independent contour boundary as boundary morphology information to form a set of candidate regions for a single micropore.

[0007] In one optional implementation, a multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. The steps of fusing shallow geometric features and deep semantic features across layers to obtain the fused features include: Each micropore candidate region in the single micropore candidate region set is input into a deep convolutional network. Shallow and deep feature maps are extracted through multi-layer convolution operations. The shallow feature map retains the edge contour and local shape information of the micropore as shallow geometric features, and the deep feature map extracts defect category discrimination information as deep semantic features. The pore size range category is determined based on the area parameter in the boundary morphology information of each micropore candidate region. For different pore size range categories, a scale-adaptive pooling mechanism is used to dynamically adjust the pooling window size. The shallow feature map and the deep feature map are spatially normalized to obtain normalized shallow features and normalized deep features. The shallow fusion weights are dynamically determined based on the boundary clarity component in the shallow geometric features, and the deep fusion weights are dynamically determined based on the defect confidence component in the deep semantic features. The normalized shallow features and the normalized deep features are then weighted and fused to obtain the fused features.

[0008] In one optional implementation, the steps of dynamically determining shallow fusion weights based on the boundary sharpness component in the shallow geometric features and dynamically determining deep fusion weights based on the defect confidence component in the deep semantic features include: The gradient intensity and contour continuity index of the micropore edge are extracted from the shallow geometric features. The variance of the edge gradient intensity and the integrity of the contour continuity are calculated. The variance and integrity are converted into boundary sharpness components through nonlinear mapping. The activation response intensity of each defect type is extracted from the deep semantic features. The interval distance between the strongest activation response and the second strongest activation response is calculated. The interval distance is converted into defect confidence components through nonlinear mapping. An adaptive weight generation network is constructed, taking the boundary clarity component and the defect confidence component as inputs. The adaptive weight generation network learns the dynamic balance relationship between shallow fusion weights and deep fusion weights, and outputs the adaptively adjusted shallow fusion weights and deep fusion weights.

[0009] In one optional implementation, the step of inputting the fused features into the classification decision module to determine the defect type and obtain the defect identification result includes: In the classification decision module, the fused features are mapped to a high-dimensional feature space. In the high-dimensional feature space, discriminant hyperplanes are established as discriminant boundaries for four defect types: aperture deviation, connected holes, blockage, and incomplete holes. The position of the fused features in the high-dimensional feature space and the geometric distance between each discriminant hyperplane are calculated. Based on the geometric distance, the classification probability of the micropore candidate region belonging to the four defect types is calculated. The classification probability is normalized to obtain the category confidence of each defect type. The defect type with the highest category confidence is selected as the initial defect type determination result of the micropore candidate region. The initial defect type determination results of all micropore candidate regions in the single micropore candidate region set are summarized with the corresponding category confidence scores to obtain the defect identification results.

[0010] In one optional implementation, a micropore spatial adjacency graph is constructed based on the spatial location information of each micropore candidate region in the defect identification result. Based on the defect type distribution pattern and confidence score of the neighboring micropores, type correction is performed on micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect types are then fed back to the adjustment of the weight parameters for cross-layer fusion. Iterative optimization improves the ability of the fusion features to distinguish between defects with blurred boundaries, resulting in a defect detection report. The steps include: The spatial location information of each micropore candidate region in the defect identification result is extracted. Each micropore candidate region is used as a node. The spatial distance between nodes is calculated based on the spatial location information. Edge connections are established between node pairs whose spatial distances meet the proximity relationship determination conditions to construct a micropore spatial adjacency graph. Traverse all nodes in the micropore spatial adjacency graph, extract the confidence score of each node, and identify nodes with confidence scores below the reliability threshold as confidence conflict nodes; calculate the difference between the confidence scores of each defect type for each node, and identify nodes with a difference between the highest confidence score and the second highest confidence score less than the discrimination threshold as boundary ambiguity nodes; For the nodes with conflicting confidence levels and the nodes with ambiguous boundaries, extract their neighboring node sets, statistically analyze the frequency of occurrence and confidence score distribution of each defect type in the neighboring node sets, and verify and correct the defect type determination result of the node based on the defect type distribution pattern of the neighboring node sets to obtain the corrected defect type; The difference between the corrected defect type and the initial defect type determination result is used as a correction feedback signal. Based on the correction feedback signal, the weight parameters in cross-layer fusion are adjusted, the fusion feature expression is regenerated, and the defect type determination and type correction are repeated until the correction difference converges. The final defect type and confidence score of all micropore candidate regions are summarized after iterative optimization, and a final defect detection report is generated.

[0011] Secondly, a smart identification system for micropore defects in filter screens based on deep convolutional networks is provided, including: The first unit is used to acquire a surface image of the metal filter to be inspected, and to perform illumination normalization and noise suppression processing on the surface image to obtain a standardized image. The second unit is used to extract micropore contours from the standardized image based on morphological operators and connected component analysis rules, to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information; The third unit is used to perform multi-level feature extraction on the single micropore candidate region set through a deep convolutional network. In the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes, and the shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fourth unit is used to input the fused features into the classification decision module, determine the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generate a confidence score for each defect type to obtain the defect identification result; The fifth unit is used to construct a micropore spatial adjacency graph based on the spatial location information of each micropore candidate region in the defect identification results. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the unit corrects the type of micropore candidate regions with confidence conflicts or blurred boundaries, and feeds the corrected defect type back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the unit improves the ability of the fusion features to distinguish defects with blurred boundaries, and obtains a defect detection report.

[0012] Thirdly, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0013] This invention solves the problem of unified detection of various specifications of hydraulic filters used in different aircraft models and system parts through a scale-adaptive pooling mechanism. This allows a single system to adapt to multiple filter models, significantly improving detection versatility and economy. The cross-layer fusion technology of shallow geometric features and deep semantic features can simultaneously capture micropore edge morphological changes and semantic defect features, effectively identifying boundary ambiguities caused by hydraulic shock or fatigue. Micropore spatial adjacency analysis introduces the unique regular arrangement constraints of hydraulic filters. By considering the state of neighboring micropores to assist in judgment, it balances the requirements of high detection rate and low false positive rate, meeting aviation safety standards. The iterative optimization mechanism enables the system to continuously self-adjust based on detection results, adapting to possible changes in the filter characteristics during service. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the intelligent identification method for micropore defects in a filter screen based on a deep convolutional network, according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0016] Figure 1 This is a flowchart illustrating the intelligent identification method for micropore defects in filter screens based on deep convolutional networks according to the present invention. Figure 1 As shown, the method includes: A surface image of the metal filter screen of the aircraft hydraulic system to be inspected is acquired, and the surface image is subjected to illumination normalization and noise suppression processing to obtain a standardized image; Micropore contours are extracted from the standardized image based on morphological operators and connected component analysis rules, resulting in a set of single micropore candidate regions containing spatial location information and boundary morphology information. Multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. Furthermore, shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fused features are input into the classification decision module, which determines the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generates a confidence score for each defect type to obtain the defect identification result. Based on the spatial location information of each micropore candidate region in the defect identification results, a micropore spatial adjacency relationship graph is constructed. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the type correction is performed on the micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect type is fed back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the ability of the fusion feature to distinguish defects with blurred boundaries is improved, and a defect detection report is obtained.

[0017] In one optional implementation, the steps of acquiring a surface image of the metal filter to be inspected, and performing illumination normalization and noise suppression processing on the surface image to obtain a standardized image include: The illumination distribution pattern in the surface image is extracted based on the statistical characteristics of the gray-level histogram. The gray-level deviation between each pixel and the target illumination reference is calculated according to the illumination distribution pattern. The gray-level values ​​of each pixel in the surface image are adaptively compensated and corrected according to the gray-level deviation to eliminate regional brightness differences and obtain an illumination-normalized image. The illumination-normalized image is decomposed in the frequency domain to separate high-frequency noise components and low-frequency structural components. The high-frequency noise components are selectively suppressed based on edge-preserving constraints, while the micropore boundary information in the low-frequency structural components is preserved. The suppressed high-frequency components and the low-frequency structural components are then reconstructed and fused to obtain a normalized image.

[0018] For example, after acquiring the surface image of the metal filter screen of the aircraft hydraulic system to be inspected, the original image obtained by the industrial camera usually exhibits uneven regional brightness due to factors such as light source position offset, differences in the reflective properties of the filter screen surface, and ambient light interference. A grayscale histogram statistical analysis is performed on the original surface image, dividing the image into several spatial sub-blocks. The pixel size of each sub-block is set to 32×32 to 64×64. The mean and variance of the grayscale value distribution within each sub-block are calculated, and the illumination distribution pattern is identified by comparing the relative deviation of the mean values ​​between different sub-blocks. When a difference in the mean grayscale value between adjacent sub-blocks exceeds 20 to 30 grayscale levels, a significant regional brightness difference is determined. The peak position of the global grayscale histogram is extracted as the target illumination reference. This reference value is usually selected as the center value of the grayscale interval with the highest proportion in the histogram. For an 8-bit grayscale image, the target illumination reference generally falls within the grayscale range of 110 to 140. The grayscale deviation between the current grayscale value of each pixel and the target illumination reference is calculated. This deviation is obtained by subtracting the target reference value from the current pixel value, retaining the positive or negative sign to distinguish between overly bright and overly dark areas. Adaptive compensation correction is performed on the grayscale value of each pixel. The compensation strategy uses a non-linear mapping function, which dynamically adjusts the correction amplitude according to the absolute value of the deviation. When the absolute value of the deviation is less than 15 gray levels, the correction coefficient is set to 0.6 to 0.8; when the absolute value of the deviation is between 15 and 40 gray levels, the correction coefficient is set to 0.8 to 1.0; and when the absolute value of the deviation exceeds 40 gray levels, the correction coefficient is set to 1.0 to 1.2 to avoid over-correction that could lead to loss of image details. The compensated pixel value is the original grayscale value minus (correction coefficient × deviation), and the result is limited to the effective grayscale range of 0 to 255. Pixel values ​​exceeding the upper limit are forcibly assigned a value of 255, and pixel values ​​below the lower limit are forcibly assigned a value of 0. After pixel-by-pixel correction, an illumination-normalized image is obtained. The average brightness difference in each region of the image is reduced to within 5 to 8 gray levels, effectively eliminating detection interference caused by uneven illumination.

[0019] Frequency domain decomposition is performed on the illumination-normalized image. A two-dimensional discrete Fourier transform (DFT) is used to convert the image from the spatial domain to the frequency domain representation. The DFT is implemented using the Fast Fourier Transform (FFT) algorithm, with a computational complexity on the logarithmic order of the total number of image pixels. In the frequency domain, the image signal is decomposed into a superposition of different frequency components. Low-frequency components correspond to the overall contour and structural information of the image, while high-frequency components correspond to edge details and noise interference. A frequency boundary threshold is set to divide the frequency domain coefficients into high-frequency noise components and low-frequency structural components. This threshold is determined based on the image size and sampling rate. For a 1024×1024 pixel image, the frequency boundary is set at a position 80 to 120 frequency units away from the center of the frequency domain. The high-frequency noise component mainly includes random noise introduced by sensor thermal noise, quantization error, and high-frequency electromagnetic interference. This noise appears as high-frequency coefficients far from the center in the frequency domain. The low-frequency structural component retains the boundary contours of the micropores and the periodic texture features of the filter. These features appear as a concentrated distribution of low-frequency coefficients near the center in the frequency domain.

[0020] When selectively suppressing high-frequency noise components, an edge-preserving constraint mechanism is introduced to avoid the loss of micro-hole boundary information. The edge-preserving constraint is achieved by calculating the spatial gradient intensity of frequency domain coefficients. When a high-frequency coefficient corresponds to a pixel location in the spatial domain with an edge feature whose gradient amplitude exceeds 15 to 20, the coefficient is determined to be an edge-related component and is preserved, with its value remaining unchanged or only slightly attenuated, with the attenuation ratio set to 0.8 to 0.9. For high-frequency coefficients that do not meet the edge-preserving constraint, they are determined to be pure noise components, and a soft-threshold suppression strategy is adopted. The coefficient amplitude is subtracted from the threshold, and the positive value is taken. The threshold value is determined based on the noise intensity estimate, typically set to 2 to 3 times the noise standard deviation. The noise standard deviation is obtained through statistical analysis of the coefficients in the high-frequency region. The median absolute deviation of the coefficient amplitude within the frequency domain edge region is selected as a robust estimator, which is unaffected by local abnormally high values. The suppressed high-frequency components retain high-frequency details related to the micro-hole boundaries while effectively reducing the interference intensity of random noise.

[0021] The low-frequency structural components retain their original coefficient values ​​in the frequency domain without any suppression or attenuation, ensuring the complete preservation of the macroscopic structure and overall morphological information of the micropores. The suppressed high-frequency components are then superimposed and reconstructed with the low-frequency structural components in the frequency domain. During reconstruction, the frequency domain coefficients of the two types of components are directly summed, and then the reconstructed frequency domain representation is converted back to the spatial domain image using a two-dimensional discrete Fourier inverse transform. During the inverse transform, the real parts of the complex coefficients are taken and normalized. The normalization coefficient is set to the reciprocal of the total number of pixels in the image, ensuring that the grayscale value range of the reconstructed image remains between 0 and 255. The normalized image obtained after reconstruction and fusion maintains the clarity of the micropore boundaries while reducing the noise level to 30% to 40% of the original image, improving the image signal-to-noise ratio by 5 to 8 dB.

[0022] This method effectively preserves the details of micropore boundaries while eliminating regional brightness differences through adaptive illumination compensation and frequency-domain selective noise suppression, thereby improving the accuracy and robustness of subsequent defect identification and reducing false detection and false negative rates.

[0023] In one optional implementation, the step of extracting micropore contours from the standardized image based on morphological operators and connected component analysis rules to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information includes: The standardized image is binarized to distinguish between the micropore region and the background region. Multi-scale morphological opening and closing operations are used to fill holes and smooth edges in the binarized image to remove pseudo-connected regions caused by metal surface texture and noise residue, thus obtaining a preliminary segmented image. The preliminary segmented image is labeled with connected components, and the geometric shape parameters and topological features of each connected component are extracted. Based on the geometric shape parameters and topological features, the adhesion micropore regions are identified and separated. The boundaries of the identified adhesion micropore regions are reconstructed based on the principle of indentation point detection and watershed segmentation to determine the independent contour boundaries of each individual micropore. Traverse the independent contour boundaries, extract the center coordinates of each independent contour boundary as spatial location information, and extract the perimeter, area, roundness and major-minor axis ratio of each independent contour boundary as boundary morphology information to form a set of candidate regions for a single micropore.

[0024] For example, binarization segmentation is performed on a standardized image to accurately distinguish between micropore regions and background regions. The binarization threshold is determined using an adaptive threshold calculation method. The standardized image is divided into several local windows, each with a size set to 16×16 to 24×24 pixels. The mean and standard deviation of the pixel grayscale values ​​within each window are calculated. When the grayscale standard deviation within a window is less than 8, it is considered a uniform region, and a global threshold is applied. This global threshold is taken as the grayscale value corresponding to the valley position of the grayscale histogram of the entire image, typically falling within the grayscale range of 60 to 80. When the grayscale standard deviation within a window is greater than or equal to 8, it is considered a non-uniform region, and a local adaptive threshold is applied. This threshold is obtained by subtracting an offset of 5 to 10 from the window mean. The offset is adjusted according to the metallic reflective characteristics of the metal filter surface of the aircraft hydraulic system. For each pixel, its grayscale value is compared with the threshold at the corresponding position. Pixels with grayscale values ​​less than the threshold are identified as micropore regions and assigned a value of 0, while pixels with grayscale values ​​greater than or equal to the threshold are identified as background regions and assigned a value of 255. In the binarized segmented image, the micropores appear as black connected regions and the background appears as white regions. However, due to the influence of metal surface texture and residual noise, there are small pseudo-connected regions and burrs at the boundaries of the micropores in the binarization result.

[0025] For the binarized segmented image, morphological operations are performed on the binary image using erosion and dilation based on the shape and size of the structuring element. The structuring element uses a circular template with diameters covering three scale levels: 3 pixels, 5 pixels, and 7 pixels. Opening is performed sequentially, erosion followed by dilation. Erosion slides the structuring element across the image; when any background pixel exists within the area completely covered by the structuring element, the central pixel is set as background. This operation removes pseudo-connected regions smaller than the structuring element size and external burrs at the micro-hole boundaries. Dilation slides the structuring element across the image; when any micro-hole pixel exists within the area covered by the structuring element, the central pixel is set as a micro-hole. This operation restores the micro-hole region shrinkage caused by erosion. Closing is performed sequentially, dilation followed by erosion. Dilation fills the small holes inside the micro-holes and the internal depressions at the boundaries, while erosion restores the micro-hole region expansion caused by dilation. In multi-scale processing, one opening and one closing operation are performed for 3-pixel structuring elements, one closing operation is performed for 5-pixel structuring elements, and a closing operation is performed only for 7-pixel structuring elements when large-area holes are detected. The criteria for identifying pseudo-connected regions are those with an area less than 9 pixels or a perimeter less than 12 pixels; these regions are completely removed during morphological processing. After multi-scale morphological operations, a preliminary segmented image is obtained, in which the micropore boundaries are smooth and the interior is fully filled.

[0026] The initially segmented image is subjected to connected component labeling. Using a 4-connected or 8-connected neighborhood determination rule, all micro-hole pixels in the image are traversed, and spatially adjacent micro-hole pixels are grouped into the same connected component and assigned a unique label number. Connected component labeling is implemented using a two-pass scanning algorithm. The first pass processes each row from the top left to the bottom right, assigning a temporary label to each micro-hole pixel and establishing equivalent label pairs. The second pass merges the temporary labels into the final connected component number based on equivalence relations. Geometric parameters of each connected component are extracted, including area, perimeter, width and height of the circumscribed rectangle, and centroid coordinates. The area is obtained by counting the total number of pixels within the connected component; the perimeter is calculated by tracing the chain code length of the boundary pixels of the connected component; the circumscribed rectangle is determined by the maximum and minimum row and column coordinates of the pixels within the connected component; and the centroid coordinates are obtained by the arithmetic mean of the coordinates of all pixels within the connected component. Topological features include the Euler number of connected components and the number of concave points. The Euler number is calculated by subtracting the number of holes from the number of connected components. The number of concave points is obtained through boundary curvature analysis. The boundary curvature at a certain pixel is estimated by the cumulative angle of the three boundary pixels before and after it. When the cumulative angle exceeds 90 degrees and the direction is inward, it is determined to be a concave point.

[0027] Adhesive micropore regions are identified based on geometric parameters and topological features. The criteria for identifying adhesive micropores include a connected region area exceeding 1.6 times the standard area of ​​a single micropore, an aspect ratio of the circumscribed rectangle greater than 1.8, and a number of indentations greater than or equal to 2. The micropore diameter range of the metal filter screen in aircraft hydraulic systems is 0.05 to 0.65 mm. Under an image acquisition resolution of 8 micrometers per pixel, a 0.05 mm pore diameter corresponds to a 6-pixel diameter in the image, and a 0.65 mm pore diameter corresponds to an 81-pixel diameter. The standard area of ​​a single micropore ranges from 28 to 5150 pixels. To address the wide range of pore sizes, a segmented processing strategy is adopted for different pore size ranges. For micropores with a diameter less than 0.2 mm, the upper limit of the standard area is 490 pixels, and the adhesion threshold is set to 1.7 times the standard area, or 833 pixels. For micropores with a diameter between 0.2 and 0.4 mm, the upper limit of the standard area is 1960 pixels, and the adhesion threshold is set to 1.6 times the standard area, or 3136 pixels. For micropores with a diameter greater than 0.4 mm, the upper limit of the standard area is 5150 pixels, and the adhesion threshold is set to 1.5 times the standard area, or 7725 pixels. Identified adhered micropore regions require boundary reconstruction to separate individual micropores. Boundary reconstruction is based on the principles of indentation point detection and watershed segmentation. Indentation point detection locates local curvature maxima on the boundary curves of adhered micropores; these maxima correspond to the contact positions between the adhered micropores. Watershed segmentation uses the distance transformation map of the adhered micropore region as a topographic map. The value of each micropore pixel in the distance transformation map is the Euclidean distance from that pixel to the nearest background pixel. Local maxima in the distance transformation map correspond to the center position of each individual micropore. Using these local maxima as seed points, a flooding process is simulated, extending from low to high elevation. When water bodies from different seed points meet, a watershed line is formed, which serves as the reconstructed boundary. Indentation points are introduced as forced segmentation locations during watershed segmentation to ensure the watershed line passes through detected indentation points, avoiding over-segmentation or under-segmentation. After boundary reconstruction, the adhered micropore region is separated into multiple independent contour boundaries, each corresponding to a single micropore candidate region.

[0028] Iterate through all independent contour boundaries, extracting spatial location and boundary morphology information for each boundary. Spatial location information is obtained by calculating the mean of the row and column coordinates of all pixels within the independent contour boundary to obtain the center coordinates, expressed in pixels and rounded to one decimal place. Boundary morphology information includes perimeter, area, roundness, and the ratio of the major and minor axes. Perimeter is calculated using the pixel chain code length of the independent contour boundary. Area is calculated by counting the total number of pixels enclosed by the independent contour boundary. Roundness is calculated as (4 × pi × area) ÷ the square of the perimeter, with a value ranging from 0 to 1. The roundness of circular micropores is close to 1, while the roundness of elliptical or irregularly shaped micropores is less than 0.85. The ratio of the major and minor axes is obtained through principal component analysis (PCA) of the pixel coordinates within the independent contour boundaries. PCA calculates the covariance matrix of the pixel coordinates; the largest eigenvalue of the covariance matrix corresponds to the major axis, and the smallest eigenvalue corresponds to the minor axis. The ratio of the major and minor axes is the square root of the largest eigenvalue divided by the square root of the smallest eigenvalue. The center coordinates, perimeter, area, roundness, and major-minor axis ratio of each independent contour boundary are organized into a data structure. This data structure includes a marker number field, a center coordinate field, a perimeter field, an area field, a roundness field, and a major-minor axis ratio field. The data structures of all independent contour boundaries are summarized to form a single micropore candidate region set. This set is stored using an array or linked list and supports indexing and querying by marker number or spatial location.

[0029] This method effectively removes texture interference from metal surfaces through multi-scale morphological processing, and accurately separates adhered micropores by combining indentation detection and watershed segmentation. The extracted spatial location and morphological information provide a reliable feature basis for subsequent defect classification.

[0030] In one optional implementation, a multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. The steps of fusing shallow geometric features and deep semantic features across layers to obtain the fused features include: Each micropore candidate region in the single micropore candidate region set is input into a deep convolutional network. Shallow and deep feature maps are extracted through multi-layer convolution operations. The shallow feature map retains the edge contour and local shape information of the micropore as shallow geometric features, and the deep feature map extracts defect category discrimination information as deep semantic features. The pore size range category is determined based on the area parameter in the boundary morphology information of each micropore candidate region. For different pore size range categories, a scale-adaptive pooling mechanism is used to dynamically adjust the pooling window size. The shallow feature map and the deep feature map are spatially normalized to obtain normalized shallow features and normalized deep features. The shallow fusion weights are dynamically determined based on the boundary clarity component in the shallow geometric features, and the deep fusion weights are dynamically determined based on the defect confidence component in the deep semantic features. The normalized shallow features and the normalized deep features are then weighted and fused to obtain the fused features.

[0031] For example, each micropore candidate region in a single micropore candidate region set is sequentially input into a deep convolutional network for multi-level feature extraction. Before input, the micropore candidate regions need to undergo standardization preprocessing. During preprocessing, the bounding box size is determined based on the boundary morphology information of the micropore candidate regions. In the original standardized image, using the micropore center coordinates as a reference, the bounding box side length is 1.5 times the maximum value of the major and minor axes in the boundary morphology information, and the image patch containing the complete micropore region and the surrounding background is obtained. The image patch size dynamically changes according to the micropore diameter: the bounding box side length corresponding to micropores with diameters of 0.05 to 0.15 mm is 24 to 48 pixels, that of micropores with diameters of 0.15 to 0.35 mm is 48 to 96 pixels, and that of micropores with diameters of 0.35 to 0.65 mm is 96 to 192 pixels. The cropped image blocks are uniformly scaled to a fixed input size of 128×128 pixels using bilinear interpolation. The pixel grayscale values ​​are normalized to a floating-point range of 0 to 1. Normalization is achieved by dividing each pixel value by 255.

[0032] The deep convolutional network employs a multi-layered structure, comprising shallow and deep convolutional modules. The shallow convolutional module consists of three convolutional layers: the first layer uses 32 kernels (3×3 size, stride 1), padding of 1 pixel at the edges, and a rectified linear unit (RCU) activation function; the second layer uses 64 kernels (3×3 size, stride 1), padding of 1 pixel at the edges, and an RCU activation function; and the third layer uses 128 kernels (3×3 size, stride 1), padding of 1 pixel at the edges, and an RCU activation function. The shallow convolutional module extracts feature maps that retain the spatial resolution of the input image. The first convolutional layer outputs a feature map of size 128×128×32, the second convolutional layer outputs a feature map of size 128×128×64, and the third convolutional layer outputs a feature map of size 128×128×128. This 128×128×128 feature map serves as the shallow feature map, and its feature responses correspond to the edge contours and local shape information of the micropores, including shallow geometric features such as edge gradient direction, contour curvature changes, and corner positions. The deep convolutional module superimposes four convolutional layers and two max-pooling layers on top of the shallow convolutional module. The fourth convolutional layer uses 256 convolutional kernels with a kernel size of 3×3, a stride of 1, and a padding method of 1 pixel edge padding. The activation function is a modified linear unit. This is followed by the first max-pooling layer with a pooling window size of 2×2 and a stride of 2, reducing the feature map size to 64×64×256. The fifth convolutional layer uses 256 kernels, each with a 3×3 kernel size and a stride of 1. It uses 1-pixel padding at the edges and employs a modified linear unit (MRU) activation function. The sixth convolutional layer uses 512 kernels, each with a 3×3 kernel size and a stride of 1. It also uses 1-pixel padding at the edges and employs a MRU activation function, followed by a second max-pooling layer with a 2×2 pooling window and a stride of 2, reducing the feature map size to 32×32×512. The seventh convolutional layer uses 512 kernels, each with a 3×3 kernel size and a stride of 1. It uses 1-pixel padding at the edges and employs a MRU activation function. This 32×32×512 feature map serves as the deep feature map, where the feature responses correspond to defect category discrimination information, including semantic-level feature representations of four defect types: aperture deviation, connected holes, blockages, and incomplete holes.

[0033] The deep convolutional network was trained using supervised learning. The training dataset consisted of labeled samples of metal filters from aircraft hydraulic systems, with a total of no less than 12,000 micropore region image patches. These samples covered four defect types: aperture deviation, connected pores, blockage, and incomplete pores, as well as normal micropores. The sample proportions for each category were set as follows: aperture deviation 30%, connected pores 20%, blockage 18%, incomplete pores 18%, and normal micropores 14%. The cross-entropy loss function was used to measure the difference between the network output and the true label during training. The optimization algorithm employed adaptive moment estimation, with an initial learning rate of 0.001, which decayed to 0.5 times its original value every 20 training epochs, and a minimum learning rate of 0.00001. Batch sizes were set between 32 and 64, and the total number of training epochs was set between 100 and 150. An early stopping mechanism was triggered when the validation set loss no longer decreased for 10 consecutive epochs. The network parameters were initialized using the He Kaiming initialization method. The convolutional layer weights were randomly sampled from a normal distribution with a mean of 0 and a standard deviation dynamically determined based on the number of input channels. The bias term was initialized to 0.

[0034] Aperture range categories are determined based on the area parameters in the boundary morphology information of each micropore candidate region. The area parameter is represented by the number of pixels and converted to the actual aperture size based on the image acquisition resolution. Aperture range categories are divided into three types: areas less than 490 pixels (corresponding to apertures less than 0.2 mm) are classified as small apertures; areas between 490 and 1960 pixels (corresponding to apertures between 0.2 and 0.4 mm) are classified as medium apertures; and areas greater than 1960 pixels (corresponding to apertures greater than 0.4 mm) are classified as large apertures. A scale-adaptive pooling mechanism dynamically adjusts the pooling window size according to the aperture range category, performing adaptive pooling operations on both shallow and deep feature maps. For the small aperture category, the pooling window size for the shallow feature map is set to 4×4, pooling the 128×128 spatial dimension to obtain a normalized spatial dimension of 32×32. The pooling window size for the deep feature map is set to 1×1, maintaining the 32×32 spatial dimension unchanged. For medium-aperture categories, the pooling window size for shallow feature maps is set to 2×2, pooling the 128×128 spatial dimension to obtain a normalized spatial dimension of 64×64. For deep feature maps, the pooling window size is also set to 2×2, pooling the 32×32 spatial dimension to obtain a normalized spatial dimension of 16×16. For large-aperture categories, the pooling window size for shallow feature maps is set to 1×1, maintaining the 128×128 spatial dimension unchanged. For deep feature maps, the pooling window size is set to 4×4, pooling the 32×32 spatial dimension to obtain a normalized spatial dimension of 8×8. The pooling operation employs an adaptive average pooling method, processing each channel of the input feature map independently. The spatial dimension is divided into several sub-regions according to the pooling window size, and the arithmetic mean of the feature values ​​within each sub-region is calculated as the output value at the corresponding position after pooling. After scale-adaptive pooling, the shallow feature map is converted into normalized shallow features, and the deep feature map is converted into normalized deep features. The spatial dimensions of the two are dynamically matched according to the aperture range category, providing a consistent feature representation for subsequent cross-layer fusion.

[0035] The boundary sharpness component in shallow geometric features is obtained by analyzing the edge response intensity of normalized shallow features. Edge response intensity is obtained by calculating the spatial gradient magnitude on each channel of the normalized shallow features. The spatial gradient magnitude is calculated by taking the square root of the sum of squares of the feature value differences between adjacent pixels. The mean of the spatial gradient magnitudes of all channels in the normalized shallow features is extracted as the edge gradient intensity, with a value ranging from 0 to 5. The contour continuity index is obtained by detecting the connectivity of high-gradient response regions in the normalized shallow features. Pixels with spatial gradient magnitudes greater than 1.5 times the mean are marked as high-gradient pixels. The proportion of the largest connected region formed by high-gradient pixels to the total number of high-gradient pixels is used as the integrity of contour continuity, with a value ranging from 0 to 1. The boundary sharpness component is calculated by multiplying the edge gradient intensity and the contour continuity integrity, and then nonlinearly mapped using the hyperbolic tangent function. The input of the hyperbolic tangent function is 0.5 times the product of the edge gradient intensity and integrity, and the output range is from -1 to +1. Adding 1 to the output value and dividing by 2 yields the boundary sharpness component within the range of 0 to 1. The defect confidence component in the deep semantic features is obtained by analyzing the activation response intensity of normalized deep features to four defect types. The normalized deep features are processed by global average pooling to obtain a 512-dimensional feature vector. This feature vector is input into a fully connected layer for defect type classification. The fully connected layer contains 512 input nodes and 4 output nodes, corresponding to four defect types: aperture deviation, connected holes, blockage, and incomplete holes. The output of the fully connected layer is exponentially normalized to obtain the activation response intensity for each defect type. The activation response intensity is a probability value within the range of 0 to 1, and the sum of the probability values ​​for the four types is 1. The values ​​of the strongest and second strongest activation responses are extracted, and the difference between them is calculated as the interval distance, which ranges from 0 to 1. The defect confidence component is obtained by nonlinearly mapping the square of the interval distance. The mapped defect confidence component ranges from 0 to 1. When the difference between the strongest and second strongest activation responses is large, the defect confidence component is close to 1; when they are close, the defect confidence component is close to 0.

[0036] The adaptive weight generation network employs a two-layer fully connected structure. The input consists of two-dimensional vectors representing boundary sharpness and defect confidence components. The first fully connected layer contains 2 input nodes and 16 hidden nodes, using a modified linear unit (MRU) activation function. The second fully connected layer contains 16 input nodes and 2 output nodes, corresponding to the shallow and deep fusion weights, respectively. The output values ​​are exponentially normalized to ensure that the sum of the shallow and deep fusion weights is 1, and that each weight ranges from 0 to 1. When the boundary sharpness component is greater than 0.7, the shallow fusion weights output by the adaptive weight generation network are typically between 0.6 and 0.8, enhancing geometric shape discrimination. When the defect confidence component is greater than 0.7, the deep fusion weights are typically between 0.6 and 0.8, enhancing semantic classification. When both the boundary sharpness and defect confidence components are in the intermediate range of 0.3 to 0.7, the shallow and deep fusion weights tend to be balanced, each ranging from 0.4 to 0.6. Before weighted fusion, the normalized shallow and deep features need to have their feature dimensions unified. A 1×1 convolutional layer is used to convert both the 128 channels of the normalized shallow features and the 512 channels of the normalized deep features to 256 channels. The 1×1 convolutional layer has 256 kernels and a stride of 1, without changing the spatial dimension. The converted normalized shallow and deep features are then unified to a spatial dimension of 32×32 using bilinear interpolation. Weighted fusion is achieved through weighted summation channel-by-channel and pixel-by-pixel. The fused feature at each spatial location and channel equals the corresponding value of the normalized shallow feature multiplied by the shallow fusion weight, plus the corresponding value of the normalized deep feature multiplied by the deep fusion weight, resulting in a 32×32×256 dimensional fused feature.

[0037] This method effectively handles the feature representation of large aperture range of metal filter screens in aircraft hydraulic systems through scale-adaptive pooling mechanism. Combined with dynamic weight adjustment of boundary clarity and defect confidence, it achieves optimized fusion of shallow geometry and deep semantics, thereby improving the accuracy of defect identification.

[0038] In one optional implementation, the steps of dynamically determining shallow fusion weights based on the boundary sharpness component in the shallow geometric features and dynamically determining deep fusion weights based on the defect confidence component in the deep semantic features include: The gradient intensity and contour continuity index of the micropore edge are extracted from the shallow geometric features. The variance of the edge gradient intensity and the integrity of the contour continuity are calculated. The variance and integrity are converted into boundary sharpness components through nonlinear mapping. The activation response intensity of each defect type is extracted from the deep semantic features. The interval distance between the strongest activation response and the second strongest activation response is calculated. The interval distance is converted into defect confidence components through nonlinear mapping. An adaptive weight generation network is constructed, taking the boundary sharpness component and the defect confidence component as inputs. The adaptive weight generation network learns the dynamic balance between shallow fusion weights and deep fusion weights. When the boundary sharpness component is high, the shallow fusion weights are increased to enhance the geometric shape discrimination ability. When the defect confidence component is high, the deep fusion weights are increased to enhance the semantic classification ability. The adaptively adjusted shallow fusion weights and deep fusion weights are output.

[0039] For example, gradient intensity and contour continuity indices of micropore edges are extracted from shallow geometric features. These shallow geometric features are carried by a shallow feature map of a deep convolutional network. The spatial and channel dimensions of this feature map contain the response information of the micropore edges. Gradient intensity extraction of the micropore edges is achieved by independently calculating the spatial gradient magnitude for each channel of the shallow feature map. The spatial gradient magnitude at each pixel location is obtained by taking the square root of the sum of the squared differences in feature values ​​between that pixel and its horizontally adjacent pixels, and the squared differences in feature values ​​between that pixel and its vertically adjacent pixels. The spatial gradient magnitudes of all channels in the shallow feature map are then averaged in the spatial dimension to obtain the average gradient intensity of each channel. This average gradient intensity forms a gradient intensity vector with a length equal to the number of channels. The variance of the edge gradient intensity is obtained by dividing the sum of the squared differences between each element in the gradient intensity vector and the vector mean by the vector length. The variance ranges from 0 to positive infinity; a larger variance indicates significant differences in the intensity of the edge response between different channels, while a smaller variance indicates consistent responses across channels. The contour continuity index is obtained by analyzing the spatial distribution continuity of high-gradient response regions in the shallow feature map. The spatial gradient magnitude of each channel in the shallow feature map is maximized along the channel dimension to obtain a comprehensive gradient magnitude map for each channel. In the comprehensive gradient magnitude map, pixels with gradient magnitudes greater than 1.5 times the average gradient magnitude of the entire image are marked as high-gradient pixels. Connected component labeling is performed on these high-gradient pixels, and the proportion of high-gradient pixels contained in the largest connected region to the total number of high-gradient pixels is used as the integrity of the contour continuity. The integrity value ranges from 0 to 1. An integrity value close to 1 indicates that the micropore edge contour is continuous and concentrated, while a lower integrity value indicates that the edge contour is broken or scattered.

[0040] The variance of edge gradient intensity and the integrity of contour continuity are converted into boundary sharpness components through nonlinear mapping, which is implemented using multi-step cascaded processing. The variance value is normalized using the hyperbolic tangent function. The input to the hyperbolic tangent function is the variance value multiplied by a scaling factor, which is set between 0.1 and 0.3, with a default value of 0.2. This scaling factor adjusts the numerical range of the variance value to the sensitive range of the hyperbolic tangent function. The output range of the hyperbolic tangent function is -1 to +1. The normalized variance value in the range of 0 to 1 is obtained by dividing the output value by (output value + 1). The normalized variance value is further applied using an exponential decay function. The base of the exponential decay function is the natural constant, and the exponent is (normalized variance value - 1) × (-2). This function compresses the region where the normalized variance value is below 0.5 and stretches the region where it is above 0.5, maintaining the output range between 0 and 1. The integrity of contour continuity is enhanced through piecewise linear mapping. When the integrity is less than 0.3, the mapping output is integrity × 0.5; when the integrity is between 0.3 and 0.7, the mapping output is (integrity - 0.3) × 1.25 + 0.15; and when the integrity is greater than 0.7, the mapping output is (integrity - 0.7) × 0.75 + 0.65. This piecewise linear mapping enhances the discriminative power of regions with moderate integrity. The integrity mapping output is then smoothed using the Sigmoid function. The input to the Sigmoid function is (mapping output - 0.5) × 8, and the output is an integrity enhancement value ranging from 0 to 1. The boundary sharpness component is calculated by the geometric mean of the variance normalized value after exponential decay and the integrity enhancement value. The geometric mean is the square root of their product. This geometric mean is more sensitive to low values ​​of either component; when either component is low, the boundary sharpness component decreases significantly. The geometric mean is finally adjusted using a power function. The base of the power function is the geometric mean, and the exponent is set from 0.8 to 1.2, with a default value of 1.0. The output is the edge sharpness component, which ranges from 0 to 1. When the edge sharpness component is greater than 0.7, it is determined to be a high-definition edge, and when the edge sharpness component is less than 0.3, it is determined to be a low-definition edge.

[0041] The activation response intensities for each defect type are extracted from deep semantic features. These deep semantic features are carried by deep feature maps of a deep convolutional network, which are reduced to a one-dimensional feature vector through global average pooling. The dimension of the feature vector is equal to the number of channels in the deep feature map. The feature vector is then input into a fully connected classification layer for defect type discrimination. This layer contains the same number of input nodes as the feature vector and four output nodes, corresponding to four defect types: aperture deviation, connected holes, blockages, and incomplete holes. The output values ​​of the fully connected classification layer are converted into activation response intensities for each defect type through exponential normalization. Exponential normalization involves taking the natural exponent of each output node's value and dividing it by the sum of the exponents of all output nodes, ensuring that the sum of the activation response intensities for the four defect types is 1 and that each value ranges from 0 to 1. The magnitude of the activation response intensity reflects the confidence level of the deep semantic features in discriminating for that defect type; an activation response intensity closer to 1 indicates a higher confidence level. Calculate the strongest and second strongest activation responses among the activation response intensities of the four defect types. The strongest activation response is the largest of the four values, and the second strongest activation response is the second largest. The interval between the strongest and second strongest activation responses is obtained by subtracting the strongest activation response from the second strongest activation response. The interval ranges from 0 to 1. A larger interval indicates a high degree of distinction between the strongest category and other categories, while a smaller interval indicates category confusion.

[0042] The interval distance is converted into defect confidence components through a multi-level nonlinear mapping process, which includes three consecutive transformation stages. The interval distance undergoes initial processing via logarithmic transformation. The input to the logarithmic transformation is the interval distance + 0.01 followed by its natural logarithm. Adding 0.01 avoids the undefined logarithm when the interval distance is 0. The output range of the logarithmic transformation is from negative infinity to 0. The output value is then divided by (output value + 4.6) ÷ 4.6 to obtain a logarithmically normalized value within the range of 0 to 1, where 4.6 is the absolute value of the logarithm when the interval distance reaches its maximum value of 1.01. The logarithmically normalized value is then nonlinearly stretched using a power function. The base of the power function is the logarithmically normalized value, and the exponent is set between 1.5 and 2.5, with a default of 2.0. This exponent amplifies the differences in interval distance, making the confidence components more sensitive to changes when the interval distance is at moderate values. The power function output is compressed at its boundaries using a hyperbolic tangent function. The input to the hyperbolic tangent function is (power function output - 0.5) × 6, and the output range is from -1 to +1. The compressed value is obtained by dividing the output value by (output value + 1) ÷ 2, resulting in a compressed value within the range of 0 to 1. The compressed value is then finally adjusted using a step smoothing function, which employs a piecewise cubic polynomial. When the compressed value is less than 0.2, the output is the cube of the compressed value × 2. When the compressed value is between 0.2 and 0.8, the output is (compressed value - 0.2) × 1.67 + 0.016. When the compressed value is greater than 0.8, the output is 1 - (compressed value - 0.8) × 2. This piecewise function provides a smooth transition between low and high value regions and maintains linearity in the intermediate region. The output of the step smoothing function is the defect confidence component, ranging from 0 to 1. When the interval is 0.5, the defect confidence component is approximately 0.28; when the interval is 0.8, it is approximately 0.67; and when the interval is 1.0, it is approximately 0.98. A defect confidence component greater than 0.6 is considered a high-confidence discrimination, while a defect confidence component less than 0.3 is considered a low-confidence discrimination, requiring further analysis based on shallow geometric morphological features for auxiliary determination.

[0043] The adaptive weight generation network employs a two-layer fully connected neural network structure. The network input is a two-dimensional vector consisting of a boundary sharpness component and a defect confidence component, with both elements ranging from 0 to 1. The first fully connected layer contains 2 input nodes and 16 hidden nodes. The connection weights between nodes are obtained through offline training. The training dataset consists of labeled defect samples from metal filters in aircraft hydraulic systems, with a sample size of no less than 5000, covering feature combinations of four defect types under different aperture ranges and boundary sharpness conditions. The activation function of the first fully connected layer is a modified linear unit (MRU), which sets input values ​​less than 0 to 0 and leaves input values ​​greater than or equal to 0 unchanged. The second fully connected layer contains 16 input nodes and 2 output nodes, corresponding to the shallow fusion weights and deep fusion weights, respectively. The output values ​​of the second fully connected layer are exponentially normalized to ensure that the sum of the shallow and deep fusion weights is 1 and that each weight ranges from 0 to 1.

[0044] The training objective of the adaptive weight generation network is to minimize the cross-entropy loss of fused features in defect type discrimination. The training process employs a mini-batch gradient descent method, with batch size set to 32 to 64, learning rate set to 0.001 to 0.01, and training epochs set to 50 to 100. After training, the adaptive weight generation network exhibits the following input-output mapping characteristics: When the boundary sharpness component is greater than 0.7 and the defect confidence component is less than 0.5, the output range of shallow fusion weights is 0.65 to 0.85, and the output range of deep fusion weights is 0.15 to 0.35. This weight configuration enhances geometric shape discrimination ability to compensate for the uncertainty of semantic classification. When the defect confidence component is greater than 0.7 and the boundary sharpness component is less than 0.5, the output range of deep fusion weights is 0.65 to 0.85, and the output range of shallow fusion weights is 0.15 to 0.35. This weight configuration enhances semantic classification ability to overcome geometric shape feature distortion caused by boundary ambiguity. When both the boundary sharpness component and the defect confidence component are in the middle range of 0.4 to 0.7, the shallow fusion weights and deep fusion weights tend to be balanced, with their respective output ranges from 0.4 to 0.6, achieving collaborative discrimination of geometric shape and semantic classification. When both the boundary sharpness component and the defect confidence component are less than 0.3, the shallow fusion weights and deep fusion weights output by the adaptive weight generation network are close to 0.5, triggering a low-confidence warning mechanism, indicating that the micropore candidate region needs to be manually reviewed.

[0045] The adaptive weight generation network employs a batch processing mode for inference execution. Boundary sharpness components and defect confidence components of all micropore candidate regions within a single batch are organized into a batch input matrix. The number of rows in the batch input matrix equals the number of micropores in the batch, and the number of columns is 2. The batch input matrix undergoes matrix multiplication between the first and second fully connected layers to obtain the batch output matrix. The number of rows in the batch output matrix equals the number of micropores in the batch, and the number of columns is 2. The first column represents the shallow fusion weights, and the second column represents the deep fusion weights. The computational complexity of the batch processing mode is the number of micropores in the batch multiplied by the number of network parameters. The time taken for a single inference operation on a computing platform equipped with a graphics processing unit does not exceed 5 milliseconds, meeting the requirements for real-time detection. Network parameters are stored in single-precision floating-point format, with a total of 2×16 + 16×2 equaling 64 parameters. The parameter file size is 256 bytes, supporting fast loading and version switching.

[0046] This method refines the edge gradient variance and contour continuity through multi-level nonlinear mapping, quantifies the defect confidence by combining the cascade transformation of activation response interval distance, and uses adaptive weights to dynamically balance the fusion ratio of geometric shape and semantic features, effectively addressing the diversity and complexity of metal filter defects in aircraft hydraulic systems.

[0047] In one optional implementation, the step of inputting the fused features into the classification decision module to determine the defect type and obtain the defect identification result includes: In the classification decision module, the fused features are mapped to a high-dimensional feature space. In the high-dimensional feature space, discriminant hyperplanes are established as discriminant boundaries for four defect types: aperture deviation, connected holes, blockage, and incomplete holes. The position of the fused features in the high-dimensional feature space and the geometric distance between each discriminant hyperplane are calculated. Based on the geometric distance, the classification probability of the micropore candidate region belonging to the four defect types is calculated. The classification probability is normalized to obtain the category confidence of each defect type. The defect type with the highest category confidence is selected as the initial defect type determination result of the micropore candidate region. The initial defect type determination results of all micropore candidate regions in the single micropore candidate region set are summarized with the corresponding category confidence scores to obtain the defect identification results.

[0048] For example, the fused features are input into the classification decision module for defect type determination. The fused features are 32×32×256-dimensional three-dimensional tensors, which need to be mapped to a high-dimensional feature space to achieve linearly separable defect type discrimination. The mapping process is implemented through a fully connected layer. The number of input nodes of the fully connected layer is the total number of elements of the fused features, calculated as 32×32×256 equals 262144 nodes. The number of output nodes is set from 512 to 2048, with a default of 1024. The number of output nodes determines the dimension of the high-dimensional feature space. Before being input into the fully connected layer, the fused features need to be flattened, reorganizing the 32×32×256 three-dimensional structure into a 262144×1 one-dimensional vector. The flattening order follows the principle of spatial position priority, followed by channel dimension, that is, the 256 channel values ​​of each spatial position are arranged consecutively, and then all spatial positions are traversed in row and column order. The weight matrix of the fully connected layer is a two-dimensional matrix of 262144×1024, and the bias vector is a one-dimensional vector of 1024×1. The output of the fully connected layer is obtained by multiplying the input vector by the weight matrix and adding the bias vector, resulting in a high-dimensional feature vector of 1024×1. Each element of the high-dimensional feature vector is processed by a modified linear unit activation function. The modified linear unit sets elements less than 0 to 0, while elements greater than or equal to 0 remain unchanged. The activated high-dimensional feature vector represents the position of the fused feature in the high-dimensional feature space.

[0049] In a high-dimensional feature space, discriminative hyperplanes are established as discrimination boundaries for four defect types: aperture deviation, connected holes, blockage, and incomplete holes. These hyperplanes are obtained through support vector machine (SVM) model training. The SVM model employs a one-to-many strategy, training a binary classifier for each defect type. This binary classifier treats the current defect type as positive and the other three defect types as negative. The training dataset consists of labeled defect samples of metal filters from aircraft hydraulic systems, with a minimum of 8000 samples. The sample ratio for the four defect types is set as follows: aperture deviation 35%, connected holes 25%, blockage 20%, and incomplete holes 20%, ensuring sufficient and relatively balanced samples for each type. The kernel function of the SVM uses radial basis functions (RBFs). The bandwidth parameter of the RBF is determined through cross-validation, with a value ranging from 0.001 to 1.0, and a default value of 0.1. A penalty coefficient is used to balance the classification margin and misclassification cost, with a value ranging from 0.1 to 100, and a default value of 1.0. After training, each binary classifier corresponds to a discriminant hyperplane, which is a 1023-dimensional hypersurface in the high-dimensional feature space. The discriminant hyperplane divides the space into positive and negative regions. The discriminant hyperplanes for the four defect types are denoted as the aperture deviation discriminant hyperplane, the connected hole discriminant hyperplane, the blockage discriminant hyperplane, and the incomplete hole discriminant hyperplane.

[0050] The geometric distance between the fused feature's position in the high-dimensional feature space and each discriminant hyperplane is calculated. This geometric distance is obtained through the decision function value of the support vector machine. The decision function value is the weighted sum of the inner product of the high-dimensional feature vector and the support vectors after kernel function transformation, along with the corresponding weight coefficients, plus a bias term. The sign of the decision function value indicates which side of the discriminant hyperplane the high-dimensional feature vector lies on; a positive value indicates it lies in the positive class region, and a negative value indicates it lies in the negative class region. The absolute value of the decision function value represents the geometric distance from the high-dimensional feature vector to the discriminant hyperplane; a larger absolute value indicates a greater distance and higher classification confidence. Decision function values ​​are calculated for each of the four defect types, resulting in four geometric distance values. The geometric distance values ​​range from negative infinity to positive infinity; a larger positive value indicates higher discrimination confidence for that defect type.

[0051] The classification probabilities of candidate micropore regions belonging to four defect types are calculated based on geometric distance. The classification probabilities are mapped to a probability range of 0 to 1 using the Platt scaling method, which employs the Sigmoid function. The input to the Sigmoid function is the geometric distance multiplied by a scaling parameter plus an offset parameter. The scaling and offset parameters are obtained by fitting the posterior probabilities of the support vector machine training data. The scaling parameter ranges from 0.1 to 10, and the offset parameter ranges from -5 to 5. The output of the Sigmoid function is the classification probability within the range of 0 to 1. When the geometric distance is positive and its absolute value is large, the classification probability is close to 1; when the geometric distance is negative and its absolute value is large, the classification probability is close to 0. The classification probabilities for the four defect types are denoted as the pore size deviation classification probability, the connected pore classification probability, the blockage classification probability, and the incomplete pore classification probability, respectively. The sum of the four classification probabilities is usually not equal to 1 and requires normalization. Normalization is achieved by dividing each classification probability by the sum of the four classification probabilities. After normalization, the category confidence score for each defect type is obtained. The category confidence score ranges from 0 to 1, and the sum of the four category confidence scores equals 1. The category confidence score reflects the relative probability that a micropore candidate region belongs to each defect type. The higher the category confidence score, the stronger the reliability of the defect type determination.

[0052] The defect type with the highest category confidence score is selected as the initial defect type determination result for the micropore candidate region. This selection is achieved by comparing the numerical values ​​of the four category confidence scores. The four category confidence scores are iterated through, and the category confidence score with the largest value and its corresponding defect type index are recorded. The defect type index is an integer from 0 to 3, corresponding to aperture deviation, connected holes, blockage, and defective holes, respectively. When multiple category confidence scores are equal and all are the maximum values, the selection is based on the priority order of the defect type indexes: aperture deviation, connected holes, blockage, and defective holes. This priority order is determined based on the assessment of the hazard level of different defect types in the aircraft hydraulic system. The initial defect type determination result includes the defect type index and the corresponding category confidence score value. The defect type index identifies the determined defect category, and the category confidence score value quantifies the reliability of the determination. When the highest class confidence is less than 0.5, the judgment result is marked as low confidence and triggers subsequent neighborhood correction processing. When the highest class confidence is greater than or equal to 0.5 and less than 0.7, the judgment result is marked as medium confidence. When the highest class confidence is greater than or equal to 0.7, the judgment result is marked as high confidence.

[0053] The initial defect type determination results and corresponding category confidence scores of all micropore candidate regions in a single micropore candidate region set are summarized to obtain the defect identification results. The summarization process organizes the label number, spatial location information, defect type index, category confidence score value, and confidence status of each micropore candidate region into structured data records. Each data record corresponds to complete determination information for one micropore candidate region. The field definitions of the data records include: a label number field (integer type, ranging from 1 to the total number of micropore candidate regions); a spatial location information field containing two floating-point numbers (x-coordinate and y-coordinate of the center coordinates, ranging from the pixel coordinate range of the image); a defect type index field (integer type, ranging from 0 to 3); a category confidence score field (floating-point type, ranging from 0 to 1 and rounded to three decimal places); and a confidence status field (enumerated type, ranging from low confidence to medium confidence to high confidence). All data records are summarized into a defect identification result dataset, which is stored using an array or database table and supports sorting and filtering by label number, defect type index, or category confidence score value. The defect identification result dataset also generates statistical summary information, which includes aggregated indicators such as the number of defects detected for each defect type, the distribution ratio of each confidence level, and the average category confidence value, providing a quantitative basis for overall quality assessment and anomaly detection.

[0054] The classification decision module employs a batch processing and pipelined parallelism strategy for inference execution. The fused features within a single batch are organized into a batch input matrix, with the number of rows equal to the number of micropore candidate regions within the batch, and the number of columns being 262,144. Multiplication between the batch input matrix and the fully connected layer weight matrix utilizes matrix block computation, with a block size set to 64×64, leveraging caching to improve computational efficiency. The decision function computation for the support vector machine employs kernel function matrix pre-computation and lookup table optimization, pre-computing and caching the training set support vectors and the kernel function values ​​of the test samples to avoid redundant computation. The processing time for a single batch is no more than 20 milliseconds on a computing platform equipped with a graphics processing unit and no more than 100 milliseconds on a computing platform equipped only with a central processing unit, meeting the real-time detection requirements of industrial environments. The model parameter file includes the fully connected layer weight matrix, support vector machine support vectors, kernel function parameters, and Platt scaling parameters, with a total file size of approximately 1 to 2 megabytes, supporting fast loading and hot updates.

[0055] This method achieves accurate discrimination of four types of defects through high-dimensional feature space mapping and discriminant hyperplane separation, providing a quantitative evaluation basis for defect identification of metal filters in aircraft hydraulic systems.

[0056] In one optional implementation, a micropore spatial adjacency graph is constructed based on the spatial location information of each micropore candidate region in the defect identification result. Based on the defect type distribution pattern and confidence score of the neighboring micropores, type correction is performed on micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect types are then fed back to the adjustment of the weight parameters for cross-layer fusion. Iterative optimization improves the ability of the fusion features to distinguish between defects with blurred boundaries, resulting in a defect detection report. The steps include: The spatial location information of each micropore candidate region in the defect identification result is extracted. Each micropore candidate region is used as a node. The spatial distance between nodes is calculated based on the spatial location information. Edge connections are established between node pairs whose spatial distances meet the proximity relationship determination conditions to construct a micropore spatial adjacency graph. Traverse all nodes in the micropore spatial adjacency graph, extract the confidence score of each node, and identify nodes with confidence scores below the reliability threshold as confidence conflict nodes; calculate the difference between the confidence scores of each defect type for each node, and identify nodes with a difference between the highest confidence score and the second highest confidence score less than the discrimination threshold as boundary ambiguity nodes; For the nodes with conflicting confidence levels and the nodes with ambiguous boundaries, extract their neighboring node sets, statistically analyze the frequency of occurrence and confidence score distribution of each defect type in the neighboring node sets, and verify and correct the defect type determination result of the node based on the defect type distribution pattern of the neighboring node sets to obtain the corrected defect type; The difference between the corrected defect type and the initial defect type determination result is used as a correction feedback signal. Based on the correction feedback signal, the weight parameters in cross-layer fusion are adjusted, the fusion feature expression is regenerated, and the defect type determination and type correction are repeated until the correction difference converges. The final defect type and confidence score of all micropore candidate regions are summarized after iterative optimization, and a final defect detection report is generated.

[0057] For example, spatial location information of each micropore candidate region is extracted from the defect identification results. This spatial location information includes the center coordinates of each micropore candidate region, represented by two floating-point numbers (x-coordinate and y-coordinate), with the coordinate unit being pixels. The origin of the coordinates is located at the upper left corner of the image. A micropore spatial adjacency graph is constructed using each micropore candidate region as a node. The node attributes include a label number, center coordinates, defect type index, and category confidence value. The spatial distance between nodes is calculated to determine adjacency relationships. This spatial distance is calculated using the Euclidean distance between the center coordinates of two nodes; the calculation method is the square root of the sum of the squares of the differences in the x-coordinates and the squares of the differences in the y-coordinates. The adjacency determination criteria are based on a spatial distance threshold, which is determined according to the micropore distribution density of the metal filter screen in the aircraft hydraulic system. For filters with pore sizes ranging from 0.05 to 0.65 mm and pore counts from 30 to 35,000, the spatial distance threshold is set to 3 to 5 times the pixel size corresponding to the smallest pore size. With an image acquisition resolution of 8 micrometers per pixel, an aperture of 0.05 millimeters corresponds to a diameter of 6 pixels. The spatial distance threshold is set to 18 to 30 pixels, with a default of 24 pixels. When the spatial distance between two nodes is less than or equal to the spatial distance threshold, the two nodes are considered to be adjacent, and an edge is established between the node pairs. The edge connection is represented by an undirected edge, and the edge attributes include the label numbers of the two connected nodes and the spatial distance value. All node pairs are traversed, and the spatial distance is calculated for each pair of nodes to determine whether they satisfy the adjacent relationship. Node pairs that meet the condition are connected by an edge, and the set of edge connections constitutes the topology of the micropore spatial adjacency graph. The micropore spatial adjacency graph is stored using an adjacency list data structure. The adjacency list maintains a list of neighboring nodes for each node, storing the label numbers and spatial distance values ​​of all neighboring nodes that are connected to that node by an edge.

[0058] The process iterates through each node in the micropore spatial adjacency graph, extracting the category confidence score from the node attributes. The confidence score ranges from 0 to 1. Nodes with confidence scores below a reliability threshold are identified as confidence conflict nodes. The reliability threshold is set between 0.5 and 0.6, with a default of 0.5. This threshold is determined based on the confidence distribution statistics output by the classification decision module. Judgments below this threshold are considered unreliable and require correction using neighborhood information. The difference between the confidence scores for each defect type at each node is calculated, obtained by the difference between the highest and second-highest confidence scores. The category confidence scores for the four defect types corresponding to the node are extracted from the defect identification results. These four values ​​are sorted, and the highest value is taken as the highest confidence score, and the second-highest value as the second-highest confidence score. The difference between these two values ​​is the difference. The difference ranges from 0 to 1; a larger difference indicates a high degree of distinction between the highest and second-highest categories, while a smaller difference indicates category confusion. Nodes with a difference score less than the discrimination threshold are designated as boundary-ambiguous nodes. The discrimination threshold is set between 0.15 and 0.25, with a default value of 0.2. This threshold is determined based on the boundary characteristics of defect types and the cost of misjudgment. Nodes below this threshold are ambiguous in defect type determination. The union of confidence conflict nodes and boundary-ambiguous nodes constitutes the set of nodes that need correction, and the nodes in this set are marked as being in a state to be corrected.

[0059] For nodes in the correction state, their neighborhood node set is extracted from the adjacency list of the micropore space adjacency graph. This set contains all neighboring nodes connected to the current node by edges. The frequency of occurrence of each defect type in the neighborhood node set is calculated by traversing the set and accumulating the number of nodes corresponding to each defect type. The frequencies of occurrence for the four defect types are denoted as aperture deviation frequency, interconnected hole frequency, blockage frequency, and defective hole frequency, respectively, and the sum of these frequencies equals the total number of nodes in the neighborhood node set. The confidence score distribution for each defect type in the neighborhood node set is calculated by averaging the category confidence scores of all neighboring nodes of the same defect type. Only neighboring nodes whose initial defect type was determined to be that type are considered in the average calculation. The average confidence scores for the four defect types are denoted as average confidence scores for aperture deviation, interconnected holes, blockage, and defective holes, respectively. The defect type determination results of the node to be corrected are verified and corrected based on the defect type distribution pattern of the neighboring node set. The verification rule is implemented through a weighted comprehensive score of occurrence frequency and average confidence. The comprehensive score is calculated by multiplying the occurrence frequency by the total number of neighboring nodes by a weight coefficient of 0.6, and adding the average confidence by a weight coefficient of 0.4. This weighting configuration balances the contribution of neighborhood consistency and confidence reliability. The defect type corresponding to the maximum comprehensive score of the four defect types is taken as the corrected defect type. When the corrected defect type is consistent with the initial defect type determination result, the original determination result remains unchanged. When they are inconsistent, the defect type index of the node is updated to the corrected defect type, and this update operation is recorded as a type correction event.

[0060] The difference between the corrected defect type and the initial defect type determination is used as the correction feedback signal. This feedback signal includes the node label number where type correction occurred, the initial defect type index, the corrected defect type index, and the corresponding micropore candidate region features. The weight parameters in the cross-layer fusion are adjusted based on the correction feedback signal, using a backpropagation algorithm. The fused features of the micropore candidate regions with type correction and the corrected defect type labels are then re-input into the adaptive weight generation network for training. The training objective is to minimize the classification loss of the fused features on the corrected defect type. Training employs an online learning mode with a learning rate set to 0.1 to 0.3 times the initial training learning rate (default 0.2 times) to avoid drastic adjustments that could destabilize the model. Weight parameter updates utilize momentum gradient descent with a momentum coefficient of 0.9. The gradient calculation for each update is based on the current batch of corrected samples, and the batch size is set to the number of nodes with type correction, typically 5% to 15% of the total number of micropore candidate regions. After the weight parameters are updated, the fused feature representation is regenerated, and the cross-layer fusion operation of the deep convolutional network is re-executed for all micropore candidate regions to generate new fused features. The new fusion feature input classification decision module repeats the defect type determination to obtain new defect identification results. Based on the new defect identification results, the micropore spatial adjacency graph is constructed, confidence conflict and boundary ambiguity node identification is repeated, and neighborhood correction processing is performed to form an iterative optimization loop.

[0061] Convergence of iterative optimization is determined by monitoring the correction difference index, which is the proportion of nodes undergoing type correction in the current iteration to the total number of micropore candidate regions. When the decrease in the correction difference index is less than 0.5% for two consecutive iterations, convergence is considered achieved, and the iterative optimization terminates. The maximum number of iterations is set to 5 to 10, with a default of 8, to prevent excessive iteration and wasted computational resources. During iteration, the correction difference index, changes in the number of detected defects of each type, and the increase in average category confidence are recorded for each iteration to evaluate the iterative optimization effect. After iteration terminates, the final defect type and confidence score of all micropore candidate regions after iterative optimization are extracted. The final defect type is the defect type index of each node in the defect identification result of the last iteration, and the confidence score is the corresponding category confidence value.

[0062] The final defect type and confidence score of all micropore candidate areas are summarized to generate a final defect detection report. The report includes an overall filter quality assessment and defect distribution analysis. The overall quality assessment statistically analyzes the detection quantity and percentage of the four defect types, calculates the average confidence score and standard deviation of each defect type, and assesses the reliability and consistency of the test results. The defect distribution analysis generates a visual representation of the micropore spatial adjacency diagram. Node colors are distinguished according to defect type, node sizes are scaled according to confidence scores, and edge thickness reflects spatial distance. The report format supports a combination of structured text and charts. The text includes detection time, filter number, image size, total number of micropores, and detailed statistical information for each defect type. The charts include a pie chart of defect type distribution, a histogram of confidence score distribution, and a spatial distribution heatmap. The report also marks the locations of micropore candidate areas with confidence scores below 0.5, suggesting manual review to ensure the flawless quality inspection of metal filters in aircraft hydraulic systems.

[0063] This method captures the spatial distribution characteristics of defects through a micropore spatial adjacency graph. Combined with neighborhood consistency correction and iterative weight optimization, it effectively eliminates misjudgments of confidence conflicts and boundary ambiguities, providing highly reliable defect detection reports for metal filters in aircraft hydraulic systems and ensuring aviation safety.

[0064] Secondly, a smart identification system for micropore defects in filter screens based on deep convolutional networks is provided, including: The first unit is used to acquire a surface image of the metal filter to be inspected, and to perform illumination normalization and noise suppression processing on the surface image to obtain a standardized image. The second unit is used to extract micropore contours from the standardized image based on morphological operators and connected component analysis rules, to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information; The third unit is used to perform multi-level feature extraction on the single micropore candidate region set through a deep convolutional network. In the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes, and the shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fourth unit is used to input the fused features into the classification decision module, determine the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generate a confidence score for each defect type to obtain the defect identification result; The fifth unit is used to construct a micropore spatial adjacency graph based on the spatial location information of each micropore candidate region in the defect identification results. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the unit corrects the type of micropore candidate regions with confidence conflicts or blurred boundaries, and feeds the corrected defect type back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the unit improves the ability of the fusion features to distinguish defects with blurred boundaries, and obtains a defect detection report.

[0065] Thirdly, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

Claims

1. A method for intelligent identification of micropore defects in filter screens based on deep convolutional networks, characterized in that, include: A surface image of the metal filter to be tested is acquired, and the surface image is subjected to illumination normalization and noise suppression processing to obtain a standardized image; Micropore contours are extracted from the standardized image based on morphological operators and connected component analysis rules, resulting in a set of single micropore candidate regions containing spatial location information and boundary morphology information. Multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. Furthermore, shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fused features are input into the classification decision module, which determines the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generates a confidence score for each defect type to obtain the defect identification result. Based on the spatial location information of each micropore candidate region in the defect identification results, a micropore spatial adjacency relationship graph is constructed. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the type correction is performed on the micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect type is fed back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the ability of the fusion feature to distinguish defects with blurred boundaries is improved, and a defect detection report is obtained.

2. The method according to claim 1, characterized in that, The steps of acquiring a surface image of the metal filter to be tested, and performing illumination normalization and noise suppression processing on the surface image to obtain a standardized image include: The illumination distribution pattern in the surface image is extracted based on the statistical characteristics of the gray-level histogram. The gray-level deviation between each pixel and the target illumination reference is calculated according to the illumination distribution pattern. The gray-level values ​​of each pixel in the surface image are adaptively compensated and corrected according to the gray-level deviation to eliminate regional brightness differences and obtain an illumination-normalized image. The illumination-normalized image is decomposed in the frequency domain to separate high-frequency noise components and low-frequency structural components. The high-frequency noise components are selectively suppressed based on edge-preserving constraints, while the micropore boundary information in the low-frequency structural components is preserved. The suppressed high-frequency components and the low-frequency structural components are then reconstructed and fused to obtain a normalized image.

3. The method according to claim 1, characterized in that, The steps of extracting micropore contours from the standardized image based on morphological operators and connected component analysis rules to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information include: The standardized image is binarized to distinguish between the micropore region and the background region. Multi-scale morphological opening and closing operations are used to fill holes and smooth edges in the binarized image to remove pseudo-connected regions caused by metal surface texture and noise residue, thus obtaining a preliminary segmented image. The preliminary segmented image is labeled with connected components, and the geometric shape parameters and topological features of each connected component are extracted. Based on the geometric shape parameters and topological features, the adhesion micropore regions are identified and separated. The boundaries of the identified adhesion micropore regions are reconstructed based on the principle of indentation point detection and watershed segmentation to determine the independent contour boundaries of each individual micropore. Traverse the independent contour boundaries, extract the center coordinates of each independent contour boundary as spatial location information, and extract the perimeter, area, roundness and major-minor axis ratio of each independent contour boundary as boundary morphology information to form a set of candidate regions for a single micropore.

4. The method according to claim 1, characterized in that, Multi-level feature extraction is performed on the single micropore candidate region set using a deep convolutional network. During the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes. The steps for obtaining the fused features by cross-layer fusion of shallow geometric features and deep semantic features include: Each micropore candidate region in the single micropore candidate region set is input into a deep convolutional network. Shallow and deep feature maps are extracted through multi-layer convolution operations. The shallow feature map retains the edge contour and local shape information of the micropore as shallow geometric features, and the deep feature map extracts defect category discrimination information as deep semantic features. The pore size range category is determined based on the area parameter in the boundary morphology information of each micropore candidate region. For different pore size range categories, a scale-adaptive pooling mechanism is used to dynamically adjust the pooling window size. The shallow feature map and the deep feature map are spatially normalized to obtain normalized shallow features and normalized deep features. The shallow fusion weights are dynamically determined based on the boundary clarity component in the shallow geometric features, and the deep fusion weights are dynamically determined based on the defect confidence component in the deep semantic features. The normalized shallow features and the normalized deep features are then weighted and fused to obtain the fused features.

5. The method according to claim 4, characterized in that, The steps of dynamically determining shallow fusion weights based on the boundary clarity component in the shallow geometric features and dynamically determining deep fusion weights based on the defect confidence component in the deep semantic features include: The gradient intensity and contour continuity index of the micropore edge are extracted from the shallow geometric features. The variance of the edge gradient intensity and the integrity of the contour continuity are calculated. The variance and integrity are converted into boundary sharpness components through nonlinear mapping. The activation response intensity of each defect type is extracted from the deep semantic features. The interval distance between the strongest activation response and the second strongest activation response is calculated. The interval distance is converted into defect confidence components through nonlinear mapping. An adaptive weight generation network is constructed, taking the boundary clarity component and the defect confidence component as inputs. The adaptive weight generation network learns the dynamic balance relationship between shallow fusion weights and deep fusion weights, and outputs the adaptively adjusted shallow fusion weights and deep fusion weights.

6. The method according to claim 1, characterized in that, The steps of inputting the fused features into the classification decision module to determine the defect type and obtain the defect identification result include: In the classification decision module, the fused features are mapped to a high-dimensional feature space. In the high-dimensional feature space, discriminant hyperplanes are established as discriminant boundaries for four defect types: aperture deviation, connected holes, blockage, and incomplete holes. The position of the fused features in the high-dimensional feature space and the geometric distance between each discriminant hyperplane are calculated. Based on the geometric distance, the classification probability of the micropore candidate region belonging to the four defect types is calculated. The classification probability is normalized to obtain the category confidence of each defect type. The defect type with the highest category confidence is selected as the initial defect type determination result of the micropore candidate region. The initial defect type determination results of all micropore candidate regions in the single micropore candidate region set are summarized with the corresponding category confidence scores to obtain the defect identification results.

7. The method according to claim 6, characterized in that, Based on the spatial location information of each micropore candidate region in the defect identification results, a micropore spatial adjacency graph is constructed. Based on the defect type distribution pattern and confidence score of neighboring micropores, type correction is performed on micropore candidate regions with confidence conflicts or blurred boundaries. The corrected defect types are then fed back to the adjustment of the weight parameters for cross-layer fusion. Iterative optimization improves the ability of the fusion features to distinguish between blurred boundary defects, resulting in a defect detection report. The steps include: The spatial location information of each micropore candidate region in the defect identification result is extracted. Each micropore candidate region is used as a node. The spatial distance between nodes is calculated based on the spatial location information. Edge connections are established between node pairs whose spatial distances meet the proximity relationship determination conditions to construct a micropore spatial adjacency graph. Traverse all nodes in the micropore spatial adjacency graph, extract the confidence score of each node, and identify nodes with confidence scores below the reliability threshold as confidence conflict nodes; calculate the difference between the confidence scores of each defect type for each node, and identify nodes with a difference between the highest confidence score and the second highest confidence score less than the discrimination threshold as boundary ambiguity nodes; For the nodes with conflicting confidence levels and the nodes with ambiguous boundaries, extract their neighboring node sets, statistically analyze the frequency of occurrence and confidence score distribution of each defect type in the neighboring node sets, and verify and correct the defect type determination result of the node based on the defect type distribution pattern of the neighboring node sets to obtain the corrected defect type; The difference between the corrected defect type and the initial defect type determination result is used as a correction feedback signal. Based on the correction feedback signal, the weight parameters in cross-layer fusion are adjusted, the fusion feature expression is regenerated, and the defect type determination and type correction are repeated until the correction difference converges. The final defect type and confidence score of all micropore candidate regions are summarized after iterative optimization, and a final defect detection report is generated.

8. A smart identification system for micropore defects in filter screens based on deep convolutional networks, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire a surface image of the metal filter to be inspected, and to perform illumination normalization and noise suppression processing on the surface image to obtain a standardized image. The second unit is used to extract micropore contours from the standardized image based on morphological operators and connected component analysis rules, to obtain a set of single micropore candidate regions containing spatial location information and boundary morphology information; The third unit is used to perform multi-level feature extraction on the single micropore candidate region set through a deep convolutional network. In the feature extraction process, a scale-adaptive pooling mechanism is introduced to normalize the feature encoding of micropore regions with different pore sizes, and the shallow geometric features and deep semantic features are fused across layers to obtain fused features. The fourth unit is used to input the fused features into the classification decision module, determine the defect type of the micropore candidate region based on the discrimination boundary in the feature space, and generate a confidence score for each defect type to obtain the defect identification result; The fifth unit is used to construct a micropore spatial adjacency graph based on the spatial location information of each micropore candidate region in the defect identification results. Based on the defect type distribution pattern and confidence score of the neighboring micropores, the unit corrects the type of micropore candidate regions with confidence conflicts or blurred boundaries, and feeds the corrected defect type back to the adjustment of the weight parameters of cross-layer fusion. Through iterative optimization, the unit improves the ability of the fusion features to distinguish defects with blurred boundaries, and obtains a defect detection report.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images