A Multi-Algorithm Fusion Intelligent Defect Detection Method and System for Steel Claws Across All Dimensions
By employing a multi-algorithm fusion detection method, which combines texture feature extraction, image enhancement, and stereo vision reconstruction techniques, the problem of incomplete defect type coverage and high missed detection rate of subtle defects in steel claw detection has been solved, achieving high-precision, quantifiable, and full-dimensional intelligent detection.
Patent Information
- Application Number
- CN202611411287.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing steel claw detection technologies cannot fully cover multiple types of defects. They suffer from problems such as high rates of missed or false detections of minor defects, inability to quantify defect parameters, and poor adaptability to working conditions, making it difficult to achieve integrated and accurate detection of all types of defects in steel claws.
A multi-algorithm fusion detection method is adopted, including texture feature extraction combined with unsupervised anomaly screening and deep learning fine classification, image enhancement preprocessing, fusion of multi-directional temporal and frequency domain texture features, and binocular stereo vision reconstruction. Combined with a pre-built association rule base, dynamic scheduling detection is performed to achieve high-precision quantification of multi-dimensional defects.
It achieves integrated, high-precision, and quantifiable full-dimensional intelligent detection of various defects in steel claws, reducing the false detection rate and improving detection efficiency and result reliability.
Smart Images

Figure CN123118580A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-algorithm fusion-based intelligent detection method and system for all-dimensional defects in steel claws, belonging to the field of steel claw defect detection technology. Background Technology
[0002] The steel claw is a key component of the anode guide rod in the electrolytic aluminum anode assembly. It is subjected to high temperature, high pressure, and alternating loads over long periods, making it highly susceptible to various defects such as weld spalling, surface cracks, crossbeam cracks, and claw head deformation. If these defects are not detected and eliminated in a timely manner, they can lead to major safety accidents such as steel claw failure under stress, equipment leakage, and structural collapse, directly impacting industrial production safety and efficiency.
[0003] Currently, defect detection technologies for steel claws are mainly divided into three categories: The first category is the traditional manual inspection mode, in which inspectors rely on their operational experience to determine whether there are defects in the steel claws through visual observation or simple measuring tools; the second category is single machine vision inspection technology, which is based on a single image processing and feature extraction algorithm, and designs corresponding detection logic to complete the identification for specific types of steel claw defects; the third category is conventional deep learning detection technology, which uses a general image classification and object detection deep learning network, and directly inputs the collected steel claw images into the network to complete defect identification.
[0004] For the three existing steel claw detection technologies mentioned above, firstly, traditional manual inspection results heavily rely on the subjective experience of the inspectors, and the inspection process cannot output quantifiable defect feature parameters, resulting in low inspection efficiency and poor result consistency. Secondly, the algorithms of single machine vision inspection technologies are only adapted to a single preset defect type and cannot simultaneously cover multiple types of steel claw defects such as texture, cracks, and 3D deformation. Thirdly, conventional deep learning detection models have not been adapted and optimized for the industrial application scenarios and defect characteristics of steel claws, have weak resistance to environmental interference, and lack sufficient accuracy in detecting subtle defects in scenarios where the steel claw surface is reflective or obscured by dirt. In summary, existing detection technologies cannot simultaneously and comprehensively cover core defect types such as explosive weld detachment, various cracks, and claw head deformation in steel claws. They suffer from high rates of missed and false detections of subtle defects, inability to quantify defect parameters, and poor adaptability to working conditions, making it difficult to achieve integrated and accurate detection of multiple types of defects across all dimensions in steel claws. Summary of the Invention
[0005] (a) Purpose of the invention
[0006] The purpose of this invention is to provide a multi-algorithm fusion intelligent detection method and system for all-dimensional defects in steel claws, in order to solve the problems of incomplete defect type coverage, high rate of missed and false detection of minor defects, inability to quantify defect parameters, and poor adaptability to complex working conditions in existing steel claw detection technologies. This invention achieves integrated, high-precision, and quantifiable detection of various defects such as explosive weld detachment, cracks, and claw head deformation in steel claws.
[0007] (II) Technical Solution
[0008] In a first aspect, the present invention provides a multi-algorithm fusion-based intelligent defect detection method for steel claws across all dimensions, comprising:
[0009] S1. For the explosive weld detachment and abnormal surface texture defects of steel claws, a cascaded method of texture feature extraction combined with unsupervised anomaly screening and deep learning fine classification is used to detect and identify the corresponding texture defects.
[0010] S2. For surface and weld crack defects of steel claws, first optimize the defect feature recognition through image enhancement preprocessing, and then perform crack target localization, pixel-level segmentation, and morphological analysis in sequence to quantify the crack size parameters.
[0011] S3. For the crossbeam crack defect of the steel claw, multi-directional temporal texture features and frequency domain features are integrated, and detection is carried out by combining the initial identification of the classifier and the purification of deep network features.
[0012] S4. To address the deformation defects of the steel claw, a three-dimensional point cloud model of the claw head is obtained through binocular stereo vision reconstruction. After registration with the standard steel claw model, point cloud geometric deviation analysis is performed to quantify the amount of claw head deformation.
[0013] S5. Integrate the multi-dimensional defect detection results that have been completed in S1-S4, classify and grade various defects, and output a comprehensive quality inspection report for the steel claw.
[0014] Preferred options also include:
[0015] A pre-constructed steel claw defect association rule library is provided, which contains association determination rules between different types of defects in the steel claw. The association determination rules characterize the physical coupling relationship between deformation defects, surface texture defects and crack defects. Each association determination rule includes at least: the associated defect type pair, the association threshold set for the quantification parameter of the preceding defect in the defect type pair, and the association conclusion that there is a corresponding defect of another type in the associated region when the quantification parameter of the preceding defect reaches the association threshold.
[0016] The detection steps S1 to S4 are executed using dynamic scheduling: by default, texture defect detection in S1 and claw deformation detection in S4 are executed first. If no defects are detected in either step, a qualified detection report is directly output. If a defect is detected and reaches the corresponding association threshold in the association rule base, the corresponding crack detection in S2 and S3 is scheduled to be performed only for the associated region.
[0017] Preferably, the association determination rules include at least one or more of the following:
[0018] When the quantitative parameters of deformation-type defects reach the first correlation threshold, crack-type defects exist in the weld / beam area at the corresponding installation location;
[0019] When the quantification parameter of surface texture defects reaches the second correlation threshold, crack-like defects exist in the corresponding side beam area.
[0020] When the quantification parameter of crack-type defects reaches the third correlation threshold, surface texture-type defects exist in the associated region.
[0021] Preferably, before integrating the detection results in S5, a correlation verification is also included:
[0022] When the detection result matches the association rule logic in the association rule base, the defect is directly classified and graded; when it does not match the association rule logic, the corresponding area image is re-sampling and the image is enhanced for secondary detection. After the verification is passed, a comprehensive detection report is output.
[0023] Preferably, in step S1,
[0024] The texture feature extraction uses the uniform mode LBP local binary mode algorithm to extract the surface texture features of the steel claw, reducing the feature dimension and retaining the core texture information.
[0025] The unsupervised anomaly screening uses qualified steel claw texture samples as the training set to train a One-Class SVM anomaly detection model, and performs texture anomaly screening on the samples to be tested.
[0026] The deep learning fine classification uses the ResNet18 deep learning classification network to perform secondary precise classification on the initially screened abnormal areas to distinguish defect types such as explosive weld detachment, surface scratches, and texture damage.
[0027] Preferably, in step S2,
[0028] The image enhancement preprocessing adopts the multi-scale adaptive Retinex algorithm to perform illumination compensation and detail enhancement on the steel claw image in order to solve the defect blurring problem caused by surface reflection and uneven illumination.
[0029] The crack target localization adopts an improved Faster R-CNN algorithm, which realizes the localization of small targets with fine cracks by optimizing the anchor frame size and feature fusion layer;
[0030] The pixel-level segmentation employs a U-Net semantic segmentation network with skip connections to preserve crack detail features.
[0031] The morphological analysis, combined with morphological dilation and erosion algorithms, removes noise and calculates the actual length and width parameters of the crack.
[0032] Preferably, in step S3,
[0033] The multi-directional temporal texture features are extracted by setting a Gabor filter bank with multi-scale and multi-directional filtering parameters, covering all extension directions of the beam crack;
[0034] The frequency domain features are extracted by wavelet transform to obtain high-frequency defect features and low-frequency background features, thereby achieving separation of defect features and background features;
[0035] The initial identification of the classifier uses an SVM classifier to perform preliminary crack identification on the fused time-domain and frequency-domain features.
[0036] The deep network feature purification method uses a deep belief network (DBN) to purify the fused features layer by layer and optimize the feature weights, thereby suppressing noise interference and improving the accuracy of fine crack recognition.
[0037] Preferably, in step S4,
[0038] The standard model registration uses the ICP iterative nearest point algorithm to iteratively optimize the point cloud registration error and achieve accurate alignment between the three-dimensional point cloud model to be tested and the standard CAD model of the steel claw.
[0039] The point cloud geometric deviation analysis detects defects such as claw head inward bending, deformation, and dimensional offset by calculating the spatial point coordinate deviation and surface curvature deviation, and quantifies the deformation amount.
[0040] Secondly, this invention provides a multi-algorithm fusion-based intelligent defect detection system for steel claws across all dimensions, comprising:
[0041] The texture defect detection module is used to identify defects such as explosive weld detachment of steel claws and abnormal surface texture by using a cascaded method of texture feature extraction combined with initial screening of anomalies and fine classification by deep learning.
[0042] The surface weld crack detection module is used to locate, segment, and quantize cracks on the surface and weld of steel claws after image enhancement preprocessing.
[0043] The crossbeam crack detection module is used to detect minute cracks in steel claw crossbeams by fusing multi-dimensional features in the time and frequency domains, followed by initial identification and deep feature purification.
[0044] The claw head deformation detection module is used to quantitatively detect deformation defects in steel claw heads after 3D point cloud reconstruction and standard model registration.
[0045] The data fusion output module is used to integrate the detection results of the above four types of defects, complete the defect classification and grading, and output a comprehensive detection report.
[0046] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned intelligent detection method for all-dimensional defects of steel claws using multi-algorithm fusion.
[0047] Fourthly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-algorithm fusion intelligent detection method for all-dimensional defects of steel claws.
[0048] (III) Beneficial Effects
[0049] This invention presents a multi-algorithm fusion-based intelligent detection method and system for all-dimensional defects in steel claws. By employing a scenario-differentiated multi-algorithm fusion detection path, it matches dedicated multi-model cascaded detection algorithms to four types of defects: explosive weld spalling, different types of cracks, and claw head deformation. Through multi-module collaboration, it completes the fusion and output of multi-dimensional detection data, effectively solving the problems of incomplete defect coverage, high missed detection rate of minor defects, inability to quantify defect parameters, and poor adaptability to working conditions in existing steel claw detection technologies. Thus, it achieves an integrated, high-precision, quantifiable, and highly robust all-dimensional intelligent detection effect for multiple types of defects in steel claws.
[0050] Furthermore, this invention incorporates the physical coupling relationships between different types of defects in steel claws into the detection process by pre-constructing a defect association rule library, providing a basis for association verification and dynamic scheduling optimization of the detection process. Through dynamic scheduling of the detection process, texture defect detection and claw deformation detection are prioritized by default. Only when no defects are detected in the aforementioned two types of detections will a qualified detection report be directly output without activating the two types of crack detection modules. When a defect is detected and the association threshold is reached, only the corresponding crack detection module is scheduled to perform targeted detection on the associated area without full image scanning, thereby significantly saving computing power and improving detection efficiency in batch detection scenarios. Through association verification logic, after summarizing the multi-dimensional detection results, they are matched with the association rule library. When the detection results contradict the physical association rules of the defects, the corresponding area image is re-sampled and a secondary detection after image enhancement is triggered. After the verification is passed, a comprehensive detection report is output, thereby using the physical coupling relationship between defects to achieve cross-verification, further reducing the false negative rate and false positive rate, improving the accuracy of defect classification and grading, and the reliability of detection results. Attached Figure Description
[0051] Figure 1 This is a flowchart illustrating a multi-algorithm fusion-based intelligent defect detection method for steel claws across all dimensions, provided in an embodiment of the present invention.
[0052] Figure 2 This is a functional block diagram of a multi-algorithm fusion intelligent defect detection method for steel claws in all dimensions, provided in an embodiment of the present invention.
[0053] Figures 3 to 7 The images shown are the results of detection using the method described in the embodiments of the present invention. Detailed Implementation
[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0055] Reference Figure 1 The diagram shown is a flowchart illustrating a multi-algorithm fusion-based intelligent detection method for all-dimensional defects in steel claws according to an embodiment of the present invention. In this embodiment, the multi-algorithm fusion-based intelligent detection method for all-dimensional defects in steel claws includes:
[0056] S1. For defects such as explosive weld detachment and abnormal surface texture of steel claws, a cascaded method combining texture feature extraction with unsupervised anomaly screening and deep learning fine classification is used to detect and identify corresponding texture defects.
[0057] It is understood that the texture feature extraction refers to extracting numerical features from the surface image of the steel claw that reflect the texture distribution pattern and local structure, as input for subsequent anomaly judgment and classification. The unsupervised anomaly initial screening refers to establishing an anomaly judgment model using only qualified samples, without relying on defect-labeled samples, to make a preliminary judgment on whether unknown samples contain anomalies. The deep learning fine classification refers to using deep neural networks to further subdivide the regions initially judged as anomalies to determine the specific type of defect.
[0058] Furthermore, the texture feature extraction employs the uniform mode LBP local binary mode algorithm to extract the surface texture features of the steel claw, reducing the feature dimension while retaining core texture information; the unsupervised anomaly screening uses qualified steel claw texture samples as the training set to train a One-Class SVM anomaly detection model, and performs initial texture anomaly screening on the test samples; the deep learning fine classification uses a ResNet18 deep learning classification network to perform secondary precise classification on the initially screened anomaly regions to distinguish defect types such as weld spalling, surface scratches, and texture damage.
[0059] In this embodiment, to address the issues of high dimensionality, abundant redundant information, and large computational burden in directly involving classification of surface texture defects in steel claws, a uniform mode LBP local binary mode algorithm is employed to extract surface texture features of the steel claws. This reduces feature dimensionality while preserving core texture information. The specific method is as follows:
[0060] Step 1: Local Binary Encoding. Using the sampling radius... Number of sampling points For example, for each pixel in the image of the steel claw surface, its grayscale value... As the center threshold, the circumferential neighborhood grayscale value of each sampling point Instead of comparing each value individually, values greater than or equal to the threshold are recorded as 1, and values less than the threshold are recorded as 0, thus obtaining the binary code of the pixel's neighborhood. The LBP value of the pixel is then obtained by summing the values according to their positions.
[0061]
[0062] Step 2: Uniform pattern filtering. Count the number of 0 to 1 transitions in the above binary code. ,when When this mode is classified as a uniform mode, it is used to characterize the stable, core texture primitives of the steel claw surface, such as edges, spots, and flat areas; when When this occurs, it is classified as a non-uniform mode. Through uniform mode mapping, the total number of LBP modes is reduced from... Compressed to species, for example when The dimensions are compressed from 256 to 59, thereby significantly reducing the feature dimension and suppressing texture pattern redundancy caused by noise.
[0063] Step 3: Feature Map Statistical Construction. Calculate the uniform pattern LBP value for each pixel of the steel claw surface image, obtain the normalized texture feature vector by block statistical histogram, remove redundant feature dimensions, and construct a standardized texture feature dataset as input for subsequent anomaly screening and fine classification.
[0064] In this embodiment, addressing the problem of scarce defective steel claw samples in industrial settings and the difficulty in constructing a complete positive and negative sample set, a One-Class SVM anomaly detection model is trained using qualified steel claw texture samples as the training set. This achieves unsupervised initial screening of anomalies relying solely on normal samples. The specific implementation is as follows:
[0065] Step 1: Training Sample Construction. Only the texture feature vectors of qualified, defect-free steel claws are collected as the training set, eliminating the need for defective samples and addressing the class imbalance problem caused by insufficient defective samples.
[0066] Step 2: High-Dimensional Space Mapping and Boundary Learning. The training samples are mapped to a high-dimensional feature space using a kernel function. In this space, an optimal boundary (hyperplane or hypersphere) is learned that encompasses the vast majority of qualified samples, ensuring that normal samples fall inside the boundary. A commonly used kernel function is the Radial Basis Function (RBF) kernel.
[0067]
[0068] Step 3: Initial Anomaly Screening. Using the texture feature vector of the steel claw sample to be tested as input, the trained One-Class SVM model calculates its decision function value. Based on the sample's position relative to the decision boundary, it is determined whether it is an anomaly.
[0069]
[0070] When a sample to be tested falls outside the decision boundary, that is, when its texture features deviate significantly from the distribution of qualified samples, it is judged as a texture abnormal sample, and defective samples suspected of texture damage or explosive welding are quickly screened out; samples judged as normal do not need to enter the subsequent fine classification, thus achieving rapid coarse filtering in the initial screening stage.
[0071] In this embodiment, to address the issue of diverse texture defect types in the initially screened abnormal regions, requiring further differentiation of specific defect categories, a ResNet18 deep learning classification network is used to perform secondary accurate classification of the initially screened abnormal regions. The specific implementation is as follows:
[0072] Step 1: Residual Feature Extraction. The anomaly region images obtained from the initial screening are input into a ResNet18 network. The network consists of convolutional layers, batch normalization, ReLU activation, and multiple stacked residual blocks. The residual units retain features from the previous layer through identity shortcut connections.
[0073]
[0074] in, For residual mapping, This is a shortcut connection that directly passes input features. This structure effectively alleviates the gradient vanishing problem in deep network training, enabling an 18-layer network to stably extract deep semantic texture features from abnormal regions.
[0075] Step 2: Defect Category Determination. The ResNet18 network outputs the category probabilities of various defects through global average pooling and fully connected layers at the end, and completes the classification in conjunction with softmax to accurately distinguish defect types such as weld spalling, surface scratches, and texture damage. The network is trained using labeled data of qualified steel claws and various known defect samples to perform secondary accurate classification of the initially screened abnormal areas, and finally outputs the category and location of texture defects, completing the refined identification of texture defects.
[0076] In this embodiment, high-resolution images of the steel claw surface are acquired, and uniform mode LBP is used to extract image texture features. Redundant feature dimensions are removed to construct a standardized texture feature dataset. A One-Class SVM anomaly detection model is trained using qualified, defect-free steel claw samples as the training set to perform initial screening of texture anomalies on the test samples, quickly identifying defective samples suspected of texture damage or weld spalling. The initially screened anomaly regions are input into a pre-trained ResNet18 deep learning classification network for fine classification, accurately distinguishing different texture defect types such as weld spalling, surface scratches, and texture peeling.
[0077] It should be noted that this step uses a three-level cascade of "LBP texture feature extraction + One-Class SVM unsupervised initial screening + ResNet18 fine classification" to first quickly filter normal samples in an unsupervised manner, and then use a deep learning network to accurately classify abnormal regions, balancing detection speed and classification accuracy, and effectively reducing the false negative rate and false positive rate of texture defects.
[0078] S2. For surface and weld crack defects of steel claws, first optimize the defect feature recognition through image enhancement preprocessing, and then perform crack target localization, pixel-level segmentation, and morphological analysis in sequence to quantify the crack size parameters.
[0079] It is understood that the image enhancement preprocessing refers to processing the steel claw image before detection to enhance the contrast between defect features such as cracks and the background, and optimize the recognition of defect features. Crack target localization refers to detecting and defining the region where the crack defect is located in the enhanced image. Pixel-level segmentation refers to classifying the located crack region pixel by pixel and extracting the complete outline of the crack. Morphological analysis refers to post-processing the segmentation results based on morphological operations to remove noise and calculate the size parameters of the crack.
[0080] Furthermore, the image enhancement preprocessing employs a multi-scale adaptive Retinex algorithm to perform illumination compensation and detail enhancement on the steel claw image, thereby resolving the defect blurring problem caused by surface reflection and uneven illumination. The specific method is as follows:
[0081] Step 1: Image decomposition and modeling. Observe the images using the steel claw. Modeled as reflection component With light component The product of, i.e.:
[0082]
[0083] Among them, the reflection component The light component reflects the inherent reflective properties of the steel claw surface, namely defects such as cracks and welds. This reflects interference factors such as uneven lighting and surface reflection. Taking the logarithm of both sides of the above model equation, we get:
[0084]
[0085] This transforms the multiplication and integral solutions in the logarithmic field into addition separation, making it easier to estimate and remove the illumination component separately.
[0086] Step 2: Multi-scale illumination component estimation. The steel claw image is convolved with wrap functions at multiple scales to estimate the illumination components at each scale. The preferred wrap function is a Gaussian wrap function, with different standard deviations set for different scales. The implementation involves two approaches: a small-scale (small standard deviation) wrap function focuses on preserving local image details, while a large-scale (large standard deviation) wrap function focuses on estimating the overall illumination distribution to balance local and global considerations.
[0087] Step 3: Reflection Component Recovery. In the logarithmic domain, the estimated logarithms of the illumination components at each scale are subtracted from the logarithm of the observed image to recover the reflection components at the corresponding scale. We obtained detailed images at various scales after illumination compensation.
[0088] Step 4: Multi-scale adaptive weighted fusion. The reflection components recovered from each scale are fused according to their weights to obtain the final enhanced image.
[0089]
[0090] Weights of each scale The system adaptively determines the illumination uniformity of the local area in the steel claw image: In local areas with severe illumination unevenness, the weight of the large-scale component representing the overall illumination distribution is increased to more fully remove reflections and illumination interference; in local areas with rich defect details, the weight of the small-scale component is increased to preserve and highlight subtle defect features such as cracks, thereby simultaneously taking into account both overall illumination compensation and local detail enhancement.
[0091] After the above processing, the interference of surface reflection and uneven lighting in the steel claw image is effectively removed, and the detailed features of defects such as cracks are highlighted, providing a high-quality input image for subsequent crack target localization.
[0092] Furthermore, the crack target localization employs an improved Faster R-CNN algorithm, which optimizes the anchor box size and feature fusion layer to achieve localization of small targets with minute cracks. The specific method is as follows:
[0093] Step 1: Multi-scale feature extraction and fusion. A convolutional neural network is used as the backbone network to extract multi-level features from the steel claw image. A feature fusion structure is introduced to fuse deep, high-semantic features with shallow, high-resolution features layer by layer. This preserves the subtle spatial details of the crack while obtaining sufficient semantic information, forming a multi-scale fused feature map that enhances the feature representation capability of small targets with subtle cracks.
[0094] Step 2: Anchor Frame Size Optimization. Considering the slender and minute morphology of the steel claw cracks, the basic dimensions and aspect ratio of the anchor frames in the Region Proposal Network (RPN) are specifically optimized: the basic dimensions of the anchor frames are reduced to accommodate the minute crack target, and slender anchor frames with large aspect ratios are added, for example, the aspect ratio can be expanded to [missing information]. , This allows the anchor frame to better fit the slender shape of the crack, improving the recall rate of small target candidate regions.
[0095] Step 3: Candidate Region Generation and Fine Localization. The multi-scale fused feature map is input into the region proposal network to generate crack candidate regions. Foreground / background classification and preliminary location regression are then performed on the candidate regions. Subsequently, ROI alignment is performed on the candidate regions, and the detection network completes crack category determination and precise bounding box regression, outputting the precise location coordinates of the crack region, thus achieving accurate localization of small targets with minute cracks.
[0096] Furthermore, the pixel-level segmentation employs a U-Net semantic segmentation network with skip connections to preserve crack detail features. The U-Net semantic segmentation network is an encoder-decoder structure that fuses shallow detail features from the encoding stage with high-level semantic features from the decoding stage through skip connections, preserving crack details and achieving accurate pixel-level segmentation.
[0097] In this embodiment, a U-Net semantic segmentation network with skip connections is used to perform pixel-level precise segmentation of the crack region obtained from prior localization. The method is as follows:
[0098] Step 1: Encoder downsampling at each level. Using the crack region image obtained from localization as input, the encoder consists of multiple convolutional layers and pooling layers (downsampling layers), extracting features at different scales step by step and gradually compressing the spatial resolution. While deepening the network, it continuously extracts high-level semantic features of the crack region to obtain multi-scale feature representation.
[0099] Step 2: Decoder upsampling at each level. The decoder restores the spatial resolution of the feature map step by step through upsampling layers, gradually restoring the encoded high-level semantic features to the same size as the input image, providing high-resolution feature maps for pixel-by-pixel classification.
[0100] Step 3: Skip Connection Feature Fusion. Skip connections are introduced between corresponding layers in the encoder and decoder. The shallow high-resolution detail features extracted during the encoding stage are concatenated and fused with the high-level semantic features of the corresponding layers during the decoding stage at the channel dimension. This allows the network to restore spatial resolution while preserving the subtle edges and contour details of the cracks, overcoming the loss of spatial details caused by continuous downsampling. The concatenated and fused features can be represented as:
[0101]
[0102] in, For encoder number Feature map of the layer This is the feature map upsampled at the corresponding level of the decoder. After pixel-by-pixel classification, a binary image of crack segmentation is output, predicting each pixel as either a crack or background, achieving pixel-level accurate segmentation and extracting the complete outline of the crack. To address the issues of foreground pixels being much lower than background pixels in the steel claw crack and class imbalance in positive and negative samples, a cross-entropy loss function with class weights or a Dice loss function is used during training to mitigate the impact of class imbalance on the segmentation of fine cracks.
[0103] Furthermore, the morphological analysis, combined with morphological dilation and erosion algorithms, removes noise and calculates the actual crack length and width parameters. The specific method is as follows:
[0104] Step 1: Morphological denoising. Divide the crack segmentation binary image. Select structural elements First, use opening operations (erosion followed by dilation) to remove isolated noise and edge burrs:
[0105]
[0106] Then, the closing operation (expansion followed by erosion) is used to fill the tiny voids inside the crack and connect the slightly fractured crack segments:
[0107]
[0108] Among them, erosion operation Shrinking the foreground region to remove fine noise, dilation operation Expand the foreground area to fill the void. After the above processing, a cracked area with smooth edges and complete connectivity is obtained.
[0109] Step 2: Pixel Physical Size Calibration. Based on the camera calibration results, determine the conversion factor between image pixels and actual physical size, such as... Millimeters per pixel (mm / pixel) is used to convert pixel-level measurements into the actual physical size of the crack.
[0110] Step 3: Crack length and width calculation. The denoised crack region is then processed into a skeletonized structure to extract the crack centerline. The actual crack length is calculated from the length of the centerline skeleton.
[0111]
[0112] in, For the crack skeleton The pixel length accumulated along the segment direction. Number of segments for the skeleton This is a conversion factor for pixel physical dimensions. Crack width is measured along the skeleton normal direction: for each pixel in the skeleton, calculate the distance between that point and the intersection points of the crack's two edges along the normal direction. The crack width at that location is obtained, and then the average width and maximum width are calculated:
[0113]
[0114] This allows for the precise output of the actual length and width parameters of the crack, providing a quantitative basis for the classification, grading, and repair of crack defects in steel claws.
[0115] In this embodiment, the multi-scale Retinex image enhancement algorithm is used to preprocess the steel claw image, decompose the image illumination component and reflection component, compensate for uneven illumination, remove surface reflection interference, and highlight the subtle features of the crack. The enhanced image is then input into the improved Faster R-CNN detection model, which accurately locates the micro-crack and weld crack areas in the image by optimizing the small target anchor box and multi-layer feature fusion. The U-Net semantic segmentation network is used to perform pixel-level accurate segmentation of the crack area, extract the complete outline of the crack, and combine morphological dilation and erosion algorithms to remove image noise and accurately calculate the actual length and width parameters of the crack.
[0116] This step, through the chain of "Retinex enhancement + Faster R-CNN localization + U-Net segmentation + morphological quantization", first solves the problem of defect blurring caused by surface reflection and uneven lighting of the steel claw, and then completes the entire process from small target localization to pixel-level segmentation and then to size quantization, so as to achieve accurate quantization of defect parameters such as crack length and width.
[0117] S3. For the crossbeam crack defect of the steel claw, multi-directional temporal texture features and frequency domain features are integrated, and detection is carried out by combining the initial identification of the classifier and the purification of deep network features.
[0118] It is understood that the multi-directional temporal texture features refer to texture features extracted along multiple directions in the spatial (temporal) domain, used to cover the various extension directions of the beam crack. The frequency domain features refer to the frequency distribution features extracted from the image through frequency domain transformation, used to distinguish between defects (high frequency) and background (low frequency). The initial classifier identification refers to using a traditional classifier to make a preliminary judgment on the cracks based on the fused features. The deep network feature purification refers to using a deep network to optimize and denoise the fused features layer by layer, improving the discriminative power of the features.
[0119] Furthermore, the multi-directional temporal texture features are extracted using a Gabor filter bank with multi-scale and multi-directional filtering parameters, covering all extension directions of the beam crack; the frequency domain features are extracted using wavelet transform to separate high-frequency defect features and low-frequency background features, achieving separation of defect features and background features; the initial identification by the classifier uses an SVM classifier to perform preliminary crack identification on the fused temporal and frequency domain features; the deep network feature purification uses a deep belief network (DBN) to perform layer-by-layer purification and feature weight optimization on the fused features, suppressing noise interference and improving the accuracy of fine crack identification.
[0120] In this embodiment, to address the problem of diverse propagation directions of beam cracks and the difficulty in fully representing texture features in a single direction, a Gabor filter bank with multi-scale and multi-directional filtering parameters is used to extract surface texture features of the beam along multiple directions in the spatial domain (temporal domain), covering all propagation directions of the beam cracks. The specific method is as follows:
[0121] Step 1: Gabor Filter Kernel Construction. A two-dimensional Gabor filter is constructed from a complex sine function modulated by a Gaussian window function, possessing both spatial localization and direction selectivity. Its kernel function is:
[0122]
[0123] in, , ; The wavelength of the sine wave (determines the scale). In terms of filtering direction, The standard deviation is the Gaussian window value. The aspect ratio of the space. This is the phase shift.
[0124] Step 2: Multi-scale, multi-directional filtering. Set multiple directional parameters. (like , , , ) and multiple scale parameters These components are combined to form a Gabor filter bank, which performs convolution filtering on the beam image in each direction and scale to extract texture responses in multiple directions such as horizontal, vertical, and diagonal, thereby covering all extension directions of the beam crack and avoiding the missed detection of cracks in a certain direction due to filter direction mismatch.
[0125] Step 3: Texture Feature Construction. The mean, variance, and energy of the filter response amplitudes at each direction and scale are statistically analyzed and concatenated to obtain a multi-directional temporal texture feature vector, which serves as the input for subsequent fusion with frequency domain features.
[0126] In this embodiment, taking advantage of the significant difference in frequency domain distribution between cracks (high frequency) and smooth background (low frequency) in the beam image, frequency distribution features are extracted through frequency domain transformation to separate defect features from background features. The specific method is as follows:
[0127] Step 1: Wavelet Frequency Domain Decomposition. Perform multi-level discrete wavelet transform on the beam image, decomposing it into low-frequency approximate components and high-frequency detail components in the horizontal, vertical, and diagonal directions. The continuous wavelet transform can be expressed as:
[0128]
[0129] in, For wavelet basis functions, For scale parameters, These are the translation parameters.
[0130] Step 2: Frequency Domain Feature Extraction and Separation. Low-frequency approximation components correspond to smooth background regions in the image, while high-frequency detail components correspond to crack edges and abrupt texture changes. Statistical features such as energy and variance are extracted from the high-frequency detail components as high-frequency defect features, and statistical features are extracted from the low-frequency approximation components as low-frequency background features. This separates the defect (high frequency) from the background (low frequency) in the frequency domain and enhances the crack defect features.
[0131] In this embodiment, an SVM classifier is used to perform preliminary identification of beam cracks based on the fused time-domain and frequency-domain features, quickly narrowing down the scope for subsequent fine-tuning. The specific method is as follows:
[0132] Step 1: Feature Fusion. The multi-directional temporal texture feature vector is concatenated and fused with the frequency domain feature vector to obtain a fused feature vector, which serves as the input to the classifier.
[0133] Step 2: Preliminary Crack Identification. The fused feature vector is input into the SVM classifier, mapped to a high-dimensional feature space via a kernel function, and then used for preliminary crack identification. The decision function is:
[0134]
[0135] in, For kernel function, For Lagrange multipliers, For sample labels, This is a bias term. Suspected beam crack samples are quickly identified through initial screening, while background samples deemed non-cracked are removed in advance, narrowing the scope for subsequent in-depth purification.
[0136] In this embodiment, to address the issues of insufficient discriminative power and susceptibility to noise interference in the fine and fuzzy crossbeam crack features after initial screening, a Deep Belief Network (DBN) is used to refine the fused features layer by layer and optimize the feature weights. The specific method is as follows:
[0137] Step 1: Layer-by-layer feature refinement in DBN. Deep belief networks (DBNs) consist of multiple stacked Restricted Boltzmann Machines (RBMs). Unsupervised pre-training is performed layer by layer on the fused features, extracting deeper and more abstract feature representations layer by layer, thus achieving layer-by-layer feature optimization and denoising. The energy function of an RBM is:
[0138]
[0139] in, The visible layer (input fused features) Hidden layer features, For inter-layer connection weights, , These are the biases for the visible and hidden layers, respectively.
[0140] Step 2: Feature Weight Optimization and Refined Recognition. After layer-by-layer pre-training, the overall network parameters are fine-tuned through backpropagation, and the feature weights of the fused features are optimized to suppress background noise interference and highlight the discriminative features of cracks. Finally, the top layer outputs the beam crack recognition result, improving the recognition accuracy of fine and blurry beam cracks.
[0141] In this embodiment, a multi-scale, multi-directional Gabor filter bank is used to filter the beam image, extracting horizontal, vertical, and diagonal multi-dimensional texture features to cover all crack propagation directions. Wavelet transform is used to perform frequency domain decomposition on the image, separating high-frequency crack defect features from low-frequency background features to enhance defect features. The texture features and frequency domain features are fused and input into an SVM classifier to complete the initial identification of beam cracks. Finally, a deep belief network (DBN) is used to refine the fused features layer by layer and optimize the weights to suppress background noise interference and improve the recognition accuracy of fine and blurry beam cracks.
[0142] It should be noted that this step, through the fusion of time-domain and frequency-domain features, and then through the step-by-step optimization of "initial identification by classifier + purification by deep network", significantly suppresses noise interference and enhances defect features, thereby greatly improving the identification accuracy of fine and fuzzy beam cracks and reducing the false negative rate.
[0143] S4. To address the deformation defects of the steel claw's claw head, a three-dimensional point cloud model of the claw head is obtained through binocular stereo vision reconstruction. After registration with the standard steel claw model, point cloud geometric deviation analysis is performed to quantify the amount of claw head deformation.
[0144] It is understood that the aforementioned binocular stereo vision reconstruction refers to using a binocular camera to acquire left and right views and parallax information of the claw head, and reconstructing a three-dimensional point cloud model of the claw head. The aforementioned standard model registration refers to spatially aligning the reconstructed point cloud model to be tested with the standard three-dimensional model of the steel claw to eliminate installation and acquisition deviations. The aforementioned point cloud geometric deviation analysis refers to comparing the geometric differences of the two registered models point by point, detecting deformation, and quantifying the amount of deformation.
[0145] Furthermore, the standard model registration adopts the ICP iterative nearest point algorithm to iteratively optimize the point cloud registration error and achieve accurate alignment between the three-dimensional point cloud model to be tested and the standard CAD model of the steel claw.
[0146] In this embodiment, to address the issue that claw deformation is a three-dimensional morphological change that is difficult to fully represent with two-dimensional images, a binocular stereo vision method is used to reconstruct the three-dimensional point cloud model of the claw. The specific method is as follows:
[0147] Step 1: Acquisition of Left and Right Views. The left and right cameras in the binocular camera system simultaneously acquire left and right views of the claw head from different perspectives, obtaining two images of the same scene, which provide input for subsequent parallax calculation and depth restoration.
[0148] Step 2: Stereo Correction and Parallax Calculation. Epipolar correction is performed on the left and right views to ensure that the image points of the same point in space lie on the same horizontal epipolar line in both images. Then, stereo matching is used to calculate the parallax of the corresponding image points in the left and right views. That is, the difference in horizontal coordinates of the same image point in two images.
[0149] Step 3: Triangulation to reconstruct the point cloud. Based on the focal length obtained from the binocular camera calibration. With baseline length Due to parallax The depth of the corresponding spatial point of each pixel is calculated using the principle of triangulation. :
[0150]
[0151] Then, by combining pixel coordinates and camera intrinsic parameters for back projection, the three-dimensional spatial coordinates of each point are obtained, and a three-dimensional dense point cloud model of the claw is reconstructed, which fully covers the three-dimensional morphological information such as the claw's inward bending, offset, and deformation.
[0152] In this embodiment, to address the issue that the spatial pose differences between the point cloud model to be tested and the standard steel claw model exist due to differences in installation posture and acquisition perspective, and that direct comparison will introduce spurious biases, the reconstructed point cloud model to be tested and the standard 3D model of the steel claw are spatially aligned and registered. The specific method is as follows:
[0153] Step 1: Coarse registration. First, the point cloud model to be tested and the standard model are coarsely aligned by initial pose estimation. This provides a good initial pose for subsequent fine registration, accelerates iterative convergence, and avoids getting trapped in local optima.
[0154] Step 2: ICP Fine Registration. The ICP Iterative Closest Point Algorithm is used to iteratively find corresponding point pairs between the point cloud to be measured and the standard model, and the optimal rotation matrix is solved by minimizing the sum of squared distances between corresponding points. With translation vector :
[0155]
[0156] in, For points in the point cloud to be tested, For the standard model and The corresponding nearest point, This corresponds to the number of point pairs. Through multiple iterations and optimizations of the point cloud registration error until convergence, precise alignment of the 3D point cloud model under test with the standard CAD model of the steel claw is achieved, eliminating the impact of installation and acquisition deviations on subsequent geometric comparison analysis.
[0157] In this embodiment, the registered point cloud model to be tested is compared with the standard model point by point to detect the deformation of the claw and quantify the amount of deformation. The specific method is as follows:
[0158] Step 1: Calculate the spatial point coordinate deviation. For each point in the registered point cloud to be measured... Calculate its closest corresponding point on the standard model. The Euclidean distance between them is taken as the single-point deviation at that point:
[0159]
[0160] Step 2: Calculation of surface curvature deviation. Calculate the surface curvature at each point for both registered models. and Calculate the curvature deviation at the corresponding point. It is used to capture subtle deformations where the coordinate deviation of a single point is not obvious, but the local curvature of the surface has changed.
[0161] Step 3: Regional Deviation Statistics and Deformation Quantification. Combining single-point deviation and surface curvature deviation, statistical analysis is performed by dividing the area into regions according to the claw head, and the average and maximum deviations of each region are calculated:
[0162]
[0163] Simultaneously, the differences in the contour dimensions of the two models are compared. Based on the deviation values and preset thresholds, defects such as inward bending, offset, deformation, and dimensional deviation of the claw head are detected, and the specific deformation values at each location are quantified, providing a quantitative basis for classifying the deformation defects of the steel claw head.
[0164] In this embodiment, the left and right views and depth information of the steel claw head are acquired by a binocular stereo vision camera to reconstruct a three-dimensional dense point cloud model of the claw head; the standard CAD three-dimensional model of the steel claw is imported, and the ICP iterative nearest point algorithm is used to accurately register and align the three-dimensional point cloud model to be tested with the standard CAD model to eliminate the influence of installation deviation and acquisition deviation; by comparing and analyzing the differences in spatial coordinates, surface curvature, and contour dimensions of the two models, and combining the point cloud single-point deviation and regional deviation calculation methods, defects such as internal bending, offset, and deformation of the claw head are accurately detected, and the specific deformation values at each position are quantified.
[0165] It should be noted that this step, through "binocular 3D reconstruction + ICP precise registration + point cloud geometric deviation analysis", upgrades the claw deformation detection from 2D image detection to 3D geometric quantitative detection. It can output the specific deformation values at each position of the claw, realize the accurate and quantifiable detection of 3D deformation defects, and make up for the shortcomings of traditional 2D detection in that it cannot quantify deformation parameters.
[0166] S5. Integrate the multi-dimensional defect detection results that have been completed in S1-S4, classify and grade various defects, and output a comprehensive quality inspection report for the steel claw.
[0167] It is understood that the integration of multi-dimensional defect detection results refers to the summarization and fusion of defect data output from various detection steps, such as texture, crack, and 3D deformation. The classification and grading refers to classifying the detected defects by type and assessing their quality level according to their severity. The comprehensive quality inspection report refers to the summarized output containing information such as defect type, defect location, quantitative parameters, and quality grading.
[0168] This step summarizes, classifies, and outputs the detection results from the aforementioned dimensions, achieving fully automated output of the entire process from defect identification, location, and segmentation to parameter quantification and defect classification, providing complete data support for steel claw quality judgment and rework.
[0169] To verify the effectiveness of the detection method in this embodiment, a batch of steel claws was used as the detection object, and the method in this embodiment was used to detect them. The detection results are shown in Table 1.
[0170] Table 1. Detection results of steel claw defects
[0171]
[0172] Some images in the table are as follows Figures 3 to 7 As shown in the figure, the beam crack, weld crack, weld interface separation, and claw deformation correspond to four types of target defects: beam crack detection, weld and surface crack detection, explosive weld detachment detection, and claw deformation detection, respectively.
[0173] Furthermore, the multi-algorithm fusion-based intelligent defect detection method for steel claws across all dimensions also includes:
[0174] A pre-constructed steel claw defect association rule library is provided, which contains association determination rules between different types of defects in the steel claw. The association determination rules characterize the physical coupling relationship between deformation defects, surface texture defects and crack defects. Each association determination rule includes at least: the associated defect type pair, the association threshold set for the quantification parameter of the preceding defect in the defect type pair, and the association conclusion that there is a corresponding defect of another type in the associated region when the quantification parameter of the preceding defect reaches the association threshold.
[0175] The detection steps S1 to S4 are executed using dynamic scheduling: by default, texture defect detection in S1 and claw deformation detection in S4 are executed first. If no defects are detected in either step, a qualified detection report is directly output. If a defect is detected and reaches the corresponding association threshold in the association rule base, the corresponding crack detection in S2 and S3 is scheduled to be performed only for the associated region.
[0176] It is understood that the defect association rule base refers to a pre-constructed set of judgment rules describing the physical coupling relationship between different types of defects in the steel claw. For example, when the claw head deformation reaches a threshold, it is usually accompanied by a crack at the corresponding location. The association judgment rule is used to characterize the physical coupling relationship between different types of defects: "when one type of defect reaches a set condition, another type of defect is likely to appear in the corresponding area." The association threshold is a quantitative parameter threshold value that triggers the association judgment. That is, only when the quantitative parameter of the preceding defect reaches this threshold value is it determined that another type of defect may exist in the associated area. The dynamic scheduling execution refers to dynamically deciding whether to start and within what range to start the subsequent detection module based on the detected results, rather than performing a full-process scan of the entire area.
[0177] Furthermore, the association determination rules include at least one or more of the following:
[0178] When the quantification parameter of deformation-type defects reaches the first correlation threshold, crack-type defects exist in the weld / beam area at the corresponding installation location. For example, when the claw head deformation reaches the first correlation threshold, such as 0.5 mm, the probability of cracks existing in the weld / beam area at the corresponding installation location reaches 92%.
[0179] When the quantification parameter of surface texture defects reaches the second correlation threshold, crack defects exist in the corresponding side beam area; for example, when the area of explosive weld detachment reaches the second correlation threshold, such as accounting for 15% of the total explosive weld area, the probability of cracks in the corresponding side beam reaches 87%.
[0180] When the quantification parameter of crack-type defects reaches the third correlation threshold, surface texture-type defects exist in the associated region. For example, when the crack length reaches the third correlation threshold, such as 10 mm, the probability of texture damage or explosive welding defects in the corresponding region reaches 81%.
[0181] It should be noted that the association rule base described in this embodiment can be pre-established based on the simulation results of the steel claw's stress and the statistics of historical defect samples. The above thresholds and probabilities are exemplary values, and should be calibrated according to the steel claw model, material, and working conditions in actual applications.
[0182] For example, after the detection process is started, S1 texture defect detection and S4 claw deformation detection are run first by default. If no defects are detected by either, a qualified inspection report is directly output without starting the S2 and S3 crack detection modules. If S4 detects that the claw deformation reaches the corresponding association threshold in the association rule library, only S2 is scheduled to perform directional crack detection on the weld / surface area at the corresponding installation position of the claw. If S1 detects that the area of explosive weld detachment reaches the corresponding association threshold, only S3 is scheduled to perform directional crack detection on the corresponding side beam area, without the need for full image scanning.
[0183] It should be noted that the combination of defect association rule base and dynamic scheduling execution can achieve the following effects: 1. By utilizing the physical coupling and association between defects, the detection results can be targeted for verification based on the patterns of associated defects, avoiding the misjudgment of isolated or accidental detection disturbances as real defects. 2. In batch inspection scenarios with a high proportion of qualified steel claws, most workpieces only need to run the two lightweight modules S1 and S4 to complete the judgment, without the need to run all four types of detection modules in parallel, thus significantly reducing the computing resources occupied by defect-free steel claws. 3. By coupling the physical rules of defects with the multi-module detection system and combining them with intelligent dynamic scheduling, a synergistic gain of improved efficiency and reduced false detection rate is achieved without reducing detection capability.
[0184] Furthermore, the multi-algorithm fusion intelligent defect detection method for steel claws also includes: correlation verification before integrating the detection results in S5.
[0185] When the detection result matches the association rule logic in the association rule base, the defect is directly classified and graded; when it does not match the association rule logic, the corresponding area image is re-sampling and the image is enhanced for secondary detection. After the verification is passed, a comprehensive detection report is output.
[0186] The association verification refers to matching the multi-dimensional detection results with the association rule base to verify whether the detection results conform to the physical association rules between defects.
[0187] For example, S4 detects that the deformation of a steel claw head reaches the first correlation threshold in the association rule base, such as 0.5 mm, while S2 detects a crack in the weld / surface area at the corresponding installation position of the claw head. The detection results conform to the association rule logic of "claw head deformation reaches the threshold → crack exists at the corresponding position", indicating that the results of each detection module corroborate each other and are consistent. At this time, the claw head deformation defect and crack defect are directly classified and graded, and the results are sent to S5 to output a comprehensive inspection report.
[0188] For example, S4 detects that the deformation of a steel claw head reaches the first correlation threshold, but S2 fails to detect a crack in the weld / surface area at the corresponding installation location of the claw head. This result contradicts the physical correlation rule of "claw head deformation reaches the threshold → crack exists at the corresponding location," indicating the possibility of missed crack detection, such as surface reflection, dirt obscuring the image, or the crack being too fine to be detected in the first detection. In this case, image resampling of the area corresponding to the claw head's installation location is triggered. The resampling image undergoes illumination compensation and detail enhancement using a multi-scale Retinex algorithm to highlight the fine crack features. Subsequently, the S2 crack detection module performs a secondary directional detection on the resampling and enhanced image. If the secondary detection detects a crack, the verification passes, and the supplementary detection results are included in the defect classification and grading before a comprehensive inspection report is output. If the secondary detection still fails to detect a crack, the actual detection results are used as the basis for the comprehensive inspection report.
[0189] It should be noted that the correlation verification utilizes the physical coupling between defects to perform consistency cross-verification of the results from each detection module. This can identify and correct missed detections in a single module caused by factors such as lighting, reflection, dirt obstruction, or minute cracks, allowing abnormal results that violate physical laws, such as "deformation exceeding the threshold but no crack detected," to be verified, thus improving the reliability of the detection results. Simultaneously, the correlation verification is integrated with the preceding correlation rule base and dynamic scheduling execution. Dynamic scheduling first determines whether to perform targeted detection based on the correlation threshold, and the correlation verification then verifies the results based on physical correlation rules. In case of anomalies, only the correlated area is resampled and re-inspected, rather than the entire image is rescanned. Therefore, while saving computational resources, the risk of missed detections is controlled within the coverage of the correlation rules, achieving a synergy between detection efficiency and detection accuracy.
[0190] Reference Figure 2 This embodiment also provides a multi-algorithm fusion intelligent defect detection system for steel claws across all dimensions, including:
[0191] The texture defect detection module is used to identify defects such as explosive weld detachment of steel claws and abnormal surface texture by using a cascaded method of texture feature extraction combined with initial screening of anomalies and fine classification by deep learning.
[0192] The surface weld crack detection module is used to locate, segment, and quantize cracks on the surface and weld of steel claws after image enhancement preprocessing.
[0193] The crossbeam crack detection module is used to detect minute cracks in steel claw crossbeams by fusing multi-dimensional features in the time and frequency domains, followed by initial identification and deep feature purification.
[0194] The claw head deformation detection module is used to quantitatively detect deformation defects in steel claw heads after 3D point cloud reconstruction and standard model registration.
[0195] The data fusion output module is used to integrate the detection results of the above four types of defects, complete the defect classification and grading, and output a comprehensive detection report.
[0196] In detail, each module in the multi-algorithm fusion intelligent detection system for all-dimensional defects of steel claws described in this embodiment of the invention employs the same technical means as the multi-algorithm fusion intelligent detection method for all-dimensional defects of steel claws described above, and can produce the same technical effect, which will not be repeated here.
[0197] This embodiment also provides a computer-readable storage medium and an electronic device, which respectively store and execute computer programs that implement the above-described detection method to complete the detection of defects in all dimensions of the steel claw.
Claims
1. A multi-algorithm fusion steel jaw full-dimension defect intelligent detection method, characterized in that, include: S1. For the explosive weld detachment and abnormal surface texture defects of steel claws, a cascaded method of texture feature extraction combined with unsupervised anomaly screening and deep learning fine classification is used to detect and identify the corresponding texture defects. S2. For surface and weld crack defects of steel claws, first optimize the defect feature recognition through image enhancement preprocessing, and then perform crack target localization, pixel-level segmentation, and morphological analysis in sequence to quantify the crack size parameters. S3. For the crossbeam crack defect of the steel claw, multi-directional temporal texture features and frequency domain features are integrated, and detection is carried out by combining the initial identification of the classifier and the purification of deep network features. S4. To address the deformation defects of the steel claw, a three-dimensional point cloud model of the claw head is obtained through binocular stereo vision reconstruction. After registration with the standard steel claw model, point cloud geometric deviation analysis is performed to quantify the amount of claw head deformation. S5. Integrate the multi-dimensional defect detection results that have been completed in S1-S4, classify and grade various defects, and output a comprehensive quality inspection report for the steel claw.
2. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 1, characterized in that, Also includes: A pre-built rule base for steel claw defects is provided, which contains rules for determining the association between different types of defects in steel claws. The association determination rule characterizes the physical coupling relationship between deformation defects, surface texture defects and crack defects. Each association determination rule includes at least: the associated defect type pair, the association threshold set for the quantification parameter of the preceding defect in the defect type pair, and the association conclusion that there is a corresponding defect of another type in the associated region when the quantification parameter of the preceding defect reaches the association threshold. The detection steps S1 to S4 are executed using dynamic scheduling: by default, texture defect detection in S1 and claw deformation detection in S4 are executed first. If no defects are detected in either step, a qualified detection report is directly output. If a defect is detected and reaches the corresponding association threshold in the association rule base, the corresponding crack detection in S2 and S3 is scheduled to be performed only for the associated region.
3. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 2, characterized in that, The association determination rules include at least one or more of the following: When the quantitative parameters of deformation-type defects reach the first correlation threshold, crack-type defects exist in the weld / beam area at the corresponding installation location; When the quantification parameter of surface texture defects reaches the second correlation threshold, crack-like defects exist in the corresponding side beam area. When the quantification parameter of crack-type defects reaches the third correlation threshold, surface texture-type defects exist in the associated region.
4. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 2, characterized in that, Before integrating the S5 test results, correlation verification is also included: When the detection result matches the association rule logic in the association rule base, the defect is directly classified and graded; when it does not match the association rule logic, the corresponding area image is re-sampling and the image is enhanced for secondary detection. After the verification is passed, a comprehensive detection report is output.
5. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 1, characterized in that, In step S1, The texture feature extraction employs the uniform mode LBP local binary mode algorithm to extract the surface texture features of the steel claw, reducing the feature dimensionality. And retain core texture information; The unsupervised anomaly screening uses qualified steel claw texture samples as the training set to train a One-Class SVM anomaly detection model, and performs texture anomaly screening on the samples to be tested. The deep learning fine classification uses the ResNet18 deep learning classification network to perform secondary precise classification on the initially screened abnormal areas to distinguish defect types such as explosive weld detachment, surface scratches, and texture damage.
6. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 1, characterized in that, In step S2, The image enhancement preprocessing adopts the multi-scale adaptive Retinex algorithm to perform illumination compensation and detail enhancement on the steel claw image in order to solve the defect blurring problem caused by surface reflection and uneven illumination. The crack target localization adopts an improved Faster R-CNN algorithm, which realizes the localization of small targets with fine cracks by optimizing the anchor frame size and feature fusion layer; The pixel-level segmentation employs a U-Net semantic segmentation network with skip connections to preserve crack detail features. The morphological analysis, combined with morphological dilation and erosion algorithms, removes noise and calculates the actual length and width parameters of the crack.
7. The intelligent detection method for all-dimensional defects in steel claws based on multi-algorithm fusion according to claim 1, characterized in that, In step S3, The multi-directional temporal texture features are extracted by setting a Gabor filter bank with multi-scale and multi-directional filtering parameters, covering all extension directions of the beam crack; The frequency domain features are extracted by wavelet transform to obtain high-frequency defect features and low-frequency background features, thereby achieving separation of defect features and background features; The initial identification of the classifier uses an SVM classifier to perform preliminary crack identification on the fused time-domain and frequency-domain features. The deep network feature purification method uses a deep belief network (DBN) to purify the fused features layer by layer and optimize the feature weights, thereby suppressing noise interference and improving the accuracy of fine crack recognition.
8. A multi-algorithm fusion intelligent defect detection system for steel claws across all dimensions, characterized in that, include: The texture defect detection module is used to identify defects such as explosive weld detachment of steel claws and abnormal surface texture by using a cascaded method of texture feature extraction combined with initial screening of anomalies and fine classification by deep learning. The surface weld crack detection module is used to locate, segment, and quantize cracks on the surface and weld of steel claws after image enhancement preprocessing. The crossbeam crack detection module is used to detect minute cracks in steel claw crossbeams by fusing multi-dimensional features in the time and frequency domains, followed by initial identification and deep feature purification. The claw head deformation detection module is used to quantitatively detect deformation defects in steel claw heads after 3D point cloud reconstruction and standard model registration. The data fusion output module is used to integrate the detection results of the above four types of defects, complete the defect classification and grading, and output a comprehensive detection report.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent detection method for all-dimensional defects of steel claws using multi-algorithm fusion as described in any one of claims 1 to 7.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent detection method for all-dimensional defects of steel claws by multi-algorithm fusion as described in any one of claims 1 to 7.
