Visual function-based steel structure weld defect identification method and system
By using AI-driven noise adaptive weighted wavelet filtering and U-Net deep segmentation network, the problem of high-precision identification of fine cracks in weld inspection of prefabricated steel structures was solved, achieving efficient identification and stability detection in complex noise environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-10
AI Technical Summary
Existing technologies for weld inspection of prefabricated steel structures struggle to adaptively filter in complex noise environments, resulting in blurred edges of minute cracks, high false positive and false negative rates, and difficulty in achieving high-precision identification.
By employing an AI-driven noise adaptive weighted wavelet filtering strategy and a U-Net deep segmentation network, combined with a CCD camera acquisition module, the system identifies minute cracks in weld images of prefabricated steel structures through noise adaptive filtering and U-Net model segmentation anomaly detection.
It achieves high-precision identification of fine cracks in complex field environments, reduces the noise false detection rate by 30% to 45%, improves crack identification accuracy by 25% to 40%, and outputs crack distribution morphology, adapting to complex construction environments.
Smart Images
Figure CN121640177A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of weld inspection technology, specifically to a method and system for identifying weld defects in steel structures based on visual function. Background Technology
[0002] With the widespread application of prefabricated steel structure buildings in high-rise, large-span, and industrial plants, the quality of welds directly determines the overall safety performance of the structure. Micro-cracks (<0.5mm) on the weld surface are one of the main hidden dangers leading to sudden failure of steel structures. Traditional manual visual inspection combined with magnetic particle and penetrant testing methods are inefficient, costly, and difficult to achieve full-coverage real-time detection.
[0003] Existing technologies employ machine vision to identify weld defects, such as fixed threshold segmentation, mean filtering, or traditional CNN networks for defect detection. However, due to complex noise interference from uneven lighting, surface corrosion, and debris obstruction at prefabricated steel structure sites, existing fixed-parameter filtering methods are difficult to adaptively adjust, resulting in blurred edges of fine cracks, high false positive and false negative rates, and generally low crack identification accuracy in practical engineering applications.
[0004] Therefore, there is an urgent need for a visual inspection method and system that can adaptively filter and accurately identify minute cracks in welds of prefabricated steel structures in complex noise environments. Summary of the Invention
[0005] This application provides a method and system for identifying weld defects in steel structures based on visual functionality. By employing an AI-driven noise adaptive weighted wavelet filtering strategy, it achieves accurate suppression of complex on-site noise. Furthermore, by combining it with a U-Net deep segmentation network, it enables accurate identification of minute crack features in weld images of prefabricated steel structures. This improves the accuracy and robustness of weld defect identification in prefabricated steel structures and can effectively solve the problems in the background technology.
[0006] To achieve the above objectives, this application provides the following technical solution:
[0007] A visual-based method for identifying weld defects in steel structures includes the following steps:
[0008] S1: Acquire images of welds in prefabricated steel structures, and identify defective areas in the welds using a CCD camera acquisition module. These defective areas contain minute crack features (<0.5mm).
[0009] S2: Extract the initial noise pixels in the defect region, and divide each initial noise pixel according to the AI noise adaptive filtering strategy to obtain multiple initial noise pixel regions; each initial noise pixel region includes multiple initial noise pixels with a defined filtering window size;
[0010] S3: For each initial noise pixel region, calculate the noise probability of each initial noise pixel in the initial noise pixel region, and select the initial noise pixel with a noise probability greater than the preset noise probability threshold as the target noise pixel. The noise probability threshold ranges from 0.7 to 0.9.
[0011] S4: Based on the filtering window size of each target noise pixel, perform weighted wavelet adaptive filtering denoising on each target noise pixel to obtain the denoised defect region. The filter kernel size is dynamically adjusted as follows: ,in: The local noise variance of the initial noise pixel region. The maximum noise variance is preset.
[0012] S5: Perform U-Net model segmentation and anomaly detection processing on the denoised defect area to obtain the crack defect identification result of the weld image.
[0013] Furthermore, the extraction of initial noise pixels in the defect region includes:
[0014] The position information of each pixel in the defect region of the weld image is obtained, and morphological erosion operation is performed on the defect region of the weld image to obtain the first weld image.
[0015] Calculate the defect area in the weld image and the absolute value of the grayscale difference between pixels at the same location in the first weld image, and use the image containing the absolute value of the grayscale difference as the second weld image;
[0016] Based on the position information of each pixel in the second weld image and the absolute value of the grayscale difference, the optimal grayscale threshold of the second weld image is calculated, and pixels with a grayscale value greater than the optimal grayscale threshold are selected from the pixels in the second weld image as initial noise pixels.
[0017] Furthermore, according to the AI noise adaptive filtering strategy, each initial noise pixel is divided to obtain multiple initial noise pixel regions, including:
[0018] According to the preset image segmentation strategy, each initial noise pixel is divided into multiple image regions;
[0019] Identify the number of initial noise pixels contained in each image region, as well as the location information of each initial noise pixel in each image region;
[0020] For each image region, based on the number of initial noise pixels contained in the image region and the position information of each initial noise pixel, the noise recognition error rate of each initial noise pixel in the image region is calculated by the CNN model.
[0021] Based on the noise recognition error rate of each initial noise pixel in each image region and the preset filtering window size selection condition, the filtering window size corresponding to each initial noise pixel in the image region is determined, and the image region with the filtering window size of all initial noise pixels determined is taken as the initial noise pixel region.
[0022] Furthermore, determining the filter window size corresponding to each initial noise pixel in the image region based on the noise recognition error rate of each initial noise pixel in each image region and the preset filter window size filtering condition includes:
[0023] For each image region, the noise recognition error rate of each initial noise pixel in the image region is normalized to obtain the filtered value of each initial noise pixel in the image region.
[0024] Based on the filter value range corresponding to the preset filter window size and the filter value of each initial noise pixel in the image area, the filter window size corresponding to each initial noise pixel in the image area is determined. The filter value range is 0.1 to 0.9, and the filter window size range is 3x3 to 7x7. The preset filter window size selection conditions include multiple threshold ranges of noise recognition error rate, and different threshold ranges correspond to different filter window sizes.
[0025] Furthermore, for each initial noise pixel region, calculating the noise probability of each initial noise pixel point in the initial noise pixel region includes:
[0026] The location information of each initial noise pixel in each initial noise pixel region is obtained, and the distribution characteristics of each initial noise pixel in each initial noise pixel region are determined based on the location information of each initial noise pixel in each initial noise pixel region. The distribution characteristics include gradient distribution information and pixel distribution information.
[0027] For each initial noise pixel region, the noise probability of each initial noise pixel in the initial noise pixel region is calculated using the Transformer model based on the distribution characteristics of each initial noise pixel in the initial noise pixel region.
[0028] Furthermore, the step of performing weighted wavelet adaptive filtering denoising on each target noise pixel according to the filtering window size of each target noise pixel to obtain the denoised defect region includes:
[0029] The noise probability of each target noise pixel is normalized to obtain the weight of the target noise pixel. Based on the gray value of each target noise pixel, the number of initial noise pixels in the initial noise pixel region to which each target noise pixel belongs, the weight of each target noise pixel, and the filtering window size of the initial noise pixel region to which each target noise pixel belongs, the wavelet transform function is used to perform weighted adaptive filtering denoising on each target noise pixel to obtain the denoised target noise pixel.
[0030] The defect region containing all the noise-reduced target pixels is defined as the noise-reduced defect region.
[0031] Furthermore, the step of performing U-Net model segmentation and anomaly detection processing on the denoised defect region to obtain the crack defect identification result of the weld image includes:
[0032] The U-Net segmentation algorithm is used to segment the denoised defect area to obtain multiple sub-crack areas. The sub-anomaly detection value of each sub-crack area is then identified to obtain the crack defect identification result of the weld image.
[0033] A vision-based steel structure weld defect identification system includes:
[0034] The image acquisition module is used to acquire images of welds in prefabricated steel structures and identify defective areas in the welds, where the defective areas contain minute crack features (<0.5mm).
[0035] The AI noise adaptive filtering module is used to extract the initial noise pixels in the defect area, and to divide and denoise the defect area according to the AI noise adaptive filtering strategy to obtain the denoised defect area.
[0036] The crack identification module is used to perform U-Net model segmentation on the denoised defect area to obtain multiple sub-crack areas, and output the identification results of crack location, length and severity level.
[0037] Compared with the prior art, the beneficial effects of this application are:
[0038] 1. This application uses an adaptive noise filtering mechanism to automatically adjust the size of the filtering window according to the noise distribution characteristics of different weld areas. Compared with the fixed window filtering method, the noise false detection rate is reduced by 30% to 45%, effectively avoiding noise interference at the weld edge.
[0039] 2. This application achieves high-precision extraction of cracks <0.5mm by combining CNN noise error rate determination with Transformer feature association modeling. Compared with the traditional grayscale threshold segmentation method, the crack recognition accuracy is improved by 25% to 40%.
[0040] 3. The weighted wavelet adaptive filtering model constructed in this application can dynamically adjust the filtering method according to the pixel weight, improve the signal-to-noise ratio of the weld image, make the weld contour clearer, and the noise residue rate after filtering is less than 5%, thereby achieving high-precision restoration of the defect contour.
[0041] 4. This application uses the U-Net segmentation model to identify abnormal pixels, which can finely divide the weld defect area and output the crack distribution pattern. The crack identification accuracy is greater than 98%, which improves the accuracy and stability of weld defect identification and can adapt to complex on-site construction environments.
[0042] 5. This application is applicable to prefabricated steel structures, stainless steel structures and aluminum alloy weld seams, and can eliminate the influence of changes in on-site lighting conditions on the recognition results, and has high engineering versatility and industrial applicability. Attached Figure Description
[0043] Figure 1 This is a flowchart of the identification method in this application;
[0044] Figure 2 This is a schematic diagram of the system module structure of this application;
[0045] Figure 3 This is a schematic diagram of the noise adaptive filtering and crack identification process of this application. Detailed Implementation
[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0047] Please see Figure 1-3 This application provides the following technical solutions:
[0048] Example 1:
[0049] A visual-based method for identifying weld defects in steel structures includes the following steps:
[0050] S1: Acquire images of welds in prefabricated steel structures and identify defective areas in the welds using a CCD camera acquisition module. These defective areas contain minute crack features (<0.5mm).
[0051] Specifically, a CCD camera with a resolution of 2048×1536 pixels was used to image the surface of the steel structure weld under an illumination intensity of 500 lux. The acquired weld images were grayscale images with a size of 1024×1024 pixels. When identifying weld defect areas, edge detection was first performed on the weld images using the Canny edge detection algorithm. The input grayscale image was first Gaussian smoothed (Gaussian kernel size 3×3, σ=1.0), then the gradient magnitude was calculated and non-maximum suppression was performed. A dual-threshold connection method was used, setting a low threshold of 50 and a high threshold of 150 (which can be adaptively adjusted by + / -20 respectively in low light or strong reflective scenes) to extract the weld edges.
[0052] Subsequently, the edge region is expanded using morphological dilation (with a 3×3 rectangular kernel as the structuring element) to obtain preliminary defect candidate regions. Then, a region growing algorithm is used to expand from the seed point at the center of the weld, where the seed point is selected as the geometric center of the weld image, and the growth condition is that the gray-level difference between adjacent pixels is less than 20. Regions containing minute crack widths (<0.5mm) (verified by calculating pixel gradient amplitudes greater than a preset threshold of 100) are designated as defect regions. The boundaries of the defect regions are defined by minimum bounding rectangles, whose width and height do not exceed 30% of the total weld image size.
[0053] S2: Extract the initial noise pixels in the defect region.
[0054] Specifically, the positional information of each pixel in the defect region of the weld image is obtained, represented by row and column numbers. Morphological erosion is performed on the defect region of the weld image using a 3×3 cross-shaped structuring element and one iteration to obtain the first weld image. Subsequently, the absolute values of the grayscale differences between pixels in the defect region of the weld image and those at the same position in the first weld image are calculated. The image containing these absolute grayscale differences is then used as the second weld image.
[0055] For the second weld image, the optimal grayscale threshold is calculated based on the positional information of each pixel and the absolute value of the grayscale difference. The Otsu method is used to automatically determine the optimal grayscale threshold. The Otsu method calculates the threshold separately on the defect region sub-image to avoid grayscale distribution shift across the entire image. The calculated threshold is used to filter pixels with an absolute grayscale difference greater than the threshold as initial noise pixels. To reduce the influence of occasional noise, a 3×3 median filter and histogram smoothing (using 256 bins) can be applied to the image before calculating the Otsu threshold. This method segments the foreground and background by maximizing the inter-class variance.
[0056] In the second weld image, pixels whose absolute value of grayscale difference is greater than the optimal grayscale value threshold are selected as initial noise pixels.
[0057] This process ensures that the initial noisy pixels only include points with abrupt grayscale changes, and the number is controlled between 5% and 10% of the number of pixels in the defective area to avoid over-extraction.
[0058] Based on the AI noise adaptive filtering strategy, each initial noise pixel is divided into multiple initial noise pixel regions.
[0059] The "AI noise adaptive filtering strategy" referred to in this application refers to a processing strategy that divides the initial noise pixel into multiple initial noise pixel regions based on the location information and number of the initial noise pixel, calculates the noise recognition error rate through a convolutional neural network to determine the size of the filtering window, and calculates the noise probability through a Transformer model to perform weighted wavelet filtering.
[0060] Specifically, based on a preset image segmentation strategy, each initial noise pixel is divided into multiple image regions. This strategy employs the K-means clustering algorithm with a K value of 5, clustering based on the pixel's position coordinates (x, y). The maximum number of clustering iterations is 50. The K value can be adaptively adjusted based on empirical values of noise density at the engineering site or contour statistics. Each cluster forms a candidate image region, and local statistics (number of pixels, mean, variance, gradient distribution) are then calculated within this region for subsequent CNN / Transformer input. The number of initial noise pixels contained in each image region and the positional information of each initial noise pixel within each image region are identified.
[0061] For each image region, based on the number N of initial noise pixels contained in the image region and the position information of each initial noise pixel, the noise recognition error rate of each initial noise pixel in the image region is calculated by a CNN model.
[0062] This CNN model is a convolutional neural network consisting of three convolutional layers (each with a kernel size of 3×3 and 16, 32, and 64 channels respectively), followed by two fully connected layers (128 and 1 neurons respectively). The activation function is ReLU, and the output is the error rate (ranging from 0 to 1). The input is a local patch of an image region (32×32 pixels, centered at the initial noise pixel), and the training dataset consists of labeled weld noise image samples (10,000 samples). The noise recognition error rate is calculated as: Error Rate = 1 - (Number of correctly classified pixels / Total number of pixels). Based on the noise recognition error rate of each initial noise pixel in each image region and the preset filtering window size selection criteria, the corresponding filtering window size for each initial noise pixel in the image region is determined.
[0063] For each image region, the noise recognition error rate of each initial noise pixel in the image region is normalized using the formula: Filter value = (Error rate - min_error rate) / (max_error rate - min_error rate), where min_error rate and max_error rate are the minimum and maximum error rates within the region, respectively. The filter value ranges from 0.1 to 0.9.
[0064] Based on the range of filter values corresponding to the preset filter window size and the filter values of each initial noise pixel in the image region, the filter window size corresponding to each initial noise pixel in the image region is determined. The preset filter window size selection criteria include a threshold range for the noise recognition error rate: if the error rate < 0.2, the filter window size is 3×3; if 0.2 ≤ error rate < 0.5, it is 5×5; if the error rate ≥ 0.5, it is 7×7. The image region with the filter window sizes of all initial noise pixels determined is used as the initial noise pixel region. The number of initial noise pixel regions is the K value of K-means clustering. This process uses CNN to quantize the error rate to ensure that the filter window adapts to the noise density.
[0065] S3: For each initial noise pixel region, calculate the noise probability of each initial noise pixel point in the initial noise pixel region.
[0066] Specifically, the position information of each initial noise pixel in each initial noise pixel region is obtained, and the distribution characteristics of each initial noise pixel in each initial noise pixel region are determined based on the position information of each initial noise pixel in each initial noise pixel region. The distribution characteristics include gradient distribution information (the horizontal and vertical gradients are calculated using the Sobel operator to obtain the amplitude and direction) and pixel distribution information (the histogram of pixels in the region is calculated with a bin number of 256).
[0067] For each initial noise pixel region, the noise probability of each initial noise pixel in the initial noise pixel region is calculated using the Transformer model based on the distribution characteristics of each initial noise pixel in the initial noise pixel region.
[0068] This Transformer model consists of six encoder layers, each with eight attention heads, an embedding dimension of 512, a feedforward network dimension of 2048, and uses GELU activation. The input is a distribution feature vector (512 dimension), and the output is a noise probability (mapped to 0-1 using a sigmoid function). The model is trained on a labeled noisy dataset (5000 samples) using cross-entropy loss, a learning rate of 0.001, and a batch size of 32. Initial noise pixels with a probability greater than a preset threshold of 0.8 are selected as target noise pixels to ensure that the noise probability reflects the uneven distribution.
[0069] S4: Based on the filtering window size of each target noise pixel, perform weighted wavelet adaptive filtering denoising on each target noise pixel to obtain the denoised defect area.
[0070] Specifically, the noise probability of each target noise pixel is normalized to obtain its weight. The normalization formula is: Weight = Noise Probability / Sum of Noise Probabilities of All Target Noise Pixels in the Region. Based on the grayscale value G of each target noise pixel, the number M of initial noise pixels in the initial noise pixel region to which each target noise pixel belongs, the weight W of each target noise pixel, and the filtering window size of the initial noise pixel region to which each target noise pixel belongs, a weighted adaptive filtering denoising process is performed on each target noise pixel using a wavelet transform function.
[0071] The wavelet transform uses the Daubechies wavelet (order 4) with a decomposition level of 2. The filtering process includes: performing wavelet decomposition on the pixels within the window and calculating the coefficient threshold. ,in: σ is the threshold, and σ is the local noise variance of the current initial noise pixel region (obtained by calculating the pixel variance within the window). For window size, The value is the natural logarithm. A soft thresholding function is applied to the coefficients (coefficient - sign(coefficient) * T if |coefficient| > T, otherwise 0), then weighted reconstruction is performed: reconstructed pixels = sum(reconstructed coefficients * W). The filter kernel size is dynamically adjusted. ,in Set to 1.5 times the global variance of the defect region. The defect region containing all denoised target noise pixels is considered as the denoised defect region. This process uses weighting to ensure that high-probability noise points receive stronger filtering.
[0072] S5: Perform U-Net model segmentation and anomaly detection processing on the denoised defect area to obtain the crack defect identification result of the weld image.
[0073] Specifically, the U-Net segmentation algorithm is used to segment the denoised defect region. The U-Net model includes a downsampling path (4 convolutional blocks, each with two 3×3 convolutional layers, with 64, 128, 256, and 512 channels respectively) and an upsampling path (corresponding to transposed convolutions). Skip connections are used for feature fusion, the activation function is ReLU, and the output layer uses the sigmoid function. The input is the denoised defect region (size adjusted to 512×512), and training is performed on a labeled crack dataset (2000 samples) using the Dice loss function, a learning rate of 0.0001, and a batch size of 16.
[0074] Multiple sub-crack regions were segmented, and each region was extracted using connected component analysis (minimum area threshold of 10 pixels). Sub-anomaly detection values for each sub-crack region were identified, including location (center coordinates), length (calculated as pixel chain length after skeletonization), and severity level (calculated based on width: 0.4 mm for level 3; width is calculated using the minimum Feret diameter within the region). These values were then output as the crack defect identification results for the weld image.
[0075] Example 2:
[0076] This embodiment provides a specific system configuration corresponding to the above embodiments.
[0077] A vision-based steel structure weld defect identification system includes:
[0078] The image acquisition module is used to acquire images of welds in prefabricated steel structures and identify defective areas in the welds, where the defective areas contain minute crack features (<0.5mm).
[0079] This module consists of a CCD camera (resolution 2048×1536), a fixed-focus lens (focal length 12mm), a ring LED light source (color temperature 5500K, light intensity adjustable from 400 to 800 lx), an image acquisition card, and an embedded processor, in order to realize the extraction method of all replaceable defect areas in step S1 of the above embodiment.
[0080] The AI noise adaptive filtering module is used to extract the initial noise pixels in the defect area, and perform segmentation and denoising according to the AI noise adaptive filtering strategy to obtain the denoised defect area.
[0081] Deployed using NVIDIA Jetson AGX Orin (or a GPU platform with equivalent or higher computing power), integrating:
[0082] The system includes a K-means clustering module, a lightweight MobileNetV2 noise false recognition rate calculation unit, a Swin-TransformerTiny noise probability calculation unit, and a weighted wavelet denoising unit (supporting db4 / sym8 / coif5, 1-3 level decomposition, and dynamic window 3×3-9×9). These components are used to implement steps S2-S4 in the above embodiments.
[0083] The crack recognition module is used to perform U-Net model segmentation on the denoised defect area to obtain multiple sub-crack areas, and output the recognition results of crack location (center coordinates), length (pixel chain length) and severity level (level 1-3).
[0084] Deploy a U-Net (or U-Net++ / DeepLabV3+) inference model on the same GPU platform. The input is a denoised defect region (512×512), and the output is the center coordinates, length (unit: mm, pixel-to-millimeter conversion factor calibrated on-site), and severity level (1-3) of each crack. The final result is uploaded to a host computer or cloud via Ethernet or a 4G / 5G module in JSON format to achieve step S5 in the above embodiment.
[0085] The three modules are connected serially via the GigE Vision protocol or ROS2 message mechanism. The input is the original weld image, and the output is the defect identification result report. The average processing time of the entire system for a single 1024×1024 weld image is ≤180ms, which meets the real-time detection requirement of prefabricated steel structure production cycle ≤3 seconds / piece.
[0086] To verify the detection capability of this invention for fine cracks (<0.5 mm), the parameters and model of the above embodiment were used to test 320 images of prefabricated steel structure welds containing artificially created cracks (0.3 mm to 1.0 mm in width) and those collected on-site. The crack width was measured using a microscope. The statistical results are shown in Table 1.
[0087] Table 1. Statistical analysis of crack detection performance for cracks of different widths
[0088] Actual width of the crack Sample size Correct number of tests Recall rate accuracy 0.3mm-0.5mm 120 117 97.5% 96.7% 0.5mm-1.0mm 200 199 99.5% 98.9%
[0089] It is worth noting that, for ease of use and reproducibility, the network structures and training parameters given in the above embodiments are preferred example parameters. For those skilled in the art, this application is not limited to these specific values. CNN, Transformer, U-Net, and other structures can be replaced with other similar networks or variants (e.g., MobileNet, ResNet, Swin-Transformer, U-Net++, DeepLabv3+, etc.), and their hyperparameters (number of layers, number of channels, learning rate, batch size, etc.) can be reasonably adjusted according to sample size, computing power, and actual engineering needs, and are not limited by the example parameters described above.
[0090] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for identifying defects in a weld of a steel structure based on visual function, characterized by, The method comprises the following steps: S1: acquiring an assembled steel structure weld image, identifying a weld defect area through a CCD camera acquisition module, and the defect area containing fine crack features; S2: extracting initial noise pixel points in the defect area, dividing and processing each initial noise pixel point according to an AI noise adaptive filtering strategy, and obtaining a plurality of initial noise pixel regions; The initial noise pixel region includes a plurality of initial noise pixel points with determined filter window sizes; S3: for each initial noise pixel region, calculating the noise probability of each initial noise pixel point in the initial noise pixel region, and screening initial noise pixel points greater than a preset noise probability threshold as target noise pixel points, the noise probability threshold ranging from 0.7 to 0.9; S4: according to the filter window size of each target noise pixel point, performing weighted wavelet adaptive filter denoising processing on each target noise pixel point to obtain a defect region after denoising, wherein the filter core size is dynamically adjusted as wherein: is the local noise variance of the initial noise pixel region before the defect region is removed, is a preset maximum noise variance; S5: performing U-Net model segmentation and abnormality detection processing on the defect region after noise removal to obtain a crack defect recognition result of the weld image.
2. The method of claim 1, wherein: The method for extracting initial noise pixel points in the defect area comprises: Obtaining position information of each pixel point in the defect area of the weld image, and performing morphological erosion operation processing on the defect area of the weld image to obtain a first weld image; Calculating the absolute values of the gray difference values between the pixel points in the same position information of the defect area of the weld image and the first weld image, and taking the image containing the absolute values of the gray difference values as a second weld image; According to the position information and the absolute values of the gray difference values of each pixel point of the second weld image, the best gray value threshold of the second weld image is calculated, and the pixel points greater than the best gray value threshold are screened out from each pixel point of the second weld image as initial noise pixel points.
3. The method of claim 1, wherein: The method for dividing and processing each initial noise pixel point according to the AI noise adaptive filtering strategy to obtain a plurality of initial noise pixel regions comprises: According to a preset image division strategy, each initial noise pixel point is divided into a plurality of image regions; Identifying the number of initial noise pixel points contained in each image region and the position information of each initial noise pixel point in each image region; For each image region, the noise recognition failure rate of each initial noise pixel point in the image region is calculated through a CNN model according to the number of initial noise pixel points contained in the image region and the position information of each initial noise pixel point; According to the noise recognition failure rate of each initial noise pixel point in each image region and a preset filter window size screening condition, the filter window size corresponding to each initial noise pixel point in the image region is determined, and the image region in which the filter window sizes of all initial noise pixel points have been determined is taken as an initial noise pixel region.
4. The method of claim 1, wherein: The method for determining the filter window size corresponding to each initial noise pixel point in the image region according to the noise recognition failure rate of each initial noise pixel point in each image region and the preset filter window size screening condition comprises: For each image region, the noise recognition failure rate of each initial noise pixel point in the image region is normalized to obtain the filter value of each initial noise pixel point in the image region; According to the filter value range corresponding to the preset filter window size and the filter value of each initial noise pixel point in the image region, the filter window size corresponding to each initial noise pixel point in the image region is determined, wherein the filter value range is 0.1-0.9, the filter window size range is 3x3-7x7, the preset filter window size screening condition includes multiple threshold intervals of noise recognition error rates, and different threshold intervals correspond to different filter window sizes.
5. The visual function based steel structure weld defect identification method of claim 1, wherein: The method comprises the following steps: Obtain the position information of each initial noise pixel point in each initial noise pixel region, and determine the distribution characteristics of each initial noise pixel point in each initial noise pixel region according to the position information of each initial noise pixel point in each initial noise pixel region, wherein the distribution characteristics include gradient distribution information and pixel distribution information. For each initial noise pixel region, the noise probability of each initial noise pixel point in the initial noise pixel region is calculated by a Transformer model according to the distribution characteristics of each initial noise pixel point in the initial noise pixel region.
6. The visual function based steel structure weld defect identification method of claim 1, wherein: The filter window size corresponding to each target noise pixel point is determined according to the filter window size corresponding to each target noise pixel point, and each target noise pixel point is subjected to weighted wavelet adaptive filter denoising processing to obtain a denoised defect region. The noise probability of each target noise pixel point is normalized to obtain the weight value of the target noise pixel point, and each target noise pixel point is subjected to weighted adaptive filter denoising processing by a wavelet transform function according to the gray value of each target noise pixel point, the number of initial noise pixel points in the initial noise pixel region to which each target noise pixel point belongs, the weight value of each target noise pixel point, and the filter window size of the initial noise pixel region to which each target noise pixel point belongs, to obtain a denoised target noise pixel point. The defect region containing all the denoised target noise pixel points is taken as the denoised defect region.
7. The visual function based steel structure weld defect identification method of claim 1, wherein: The denoised defect region is subjected to U-Net model segmentation anomaly detection processing to obtain a crack defect recognition result of the weld image. The denoised defect region is segmented by a U-Net segmentation algorithm to obtain a plurality of sub-crack regions, and a sub-anomaly detection value of each sub-crack region is recognized to obtain a crack defect recognition result of the weld image.
8. A system for identifying defects in welds of steel structures based on visual function, characterized by, The method comprises the following steps: An image acquisition module is configured to acquire an assembled steel structure weld image and recognize a weld defect region, wherein the defect region contains fine crack characteristics. An AI noise adaptive filter module is configured to extract initial noise pixel points in the defect region and divide and denoise the initial noise pixel points according to an AI noise adaptive filter strategy to obtain a denoised defect region. A crack recognition module is configured to perform U-Net model segmentation processing on the denoised defect region to obtain a plurality of sub-crack regions, and output a recognition result of a crack position, length, and severity level.