Steel-wood combined diaphragm plate welding seam quality detection method based on visual analysis

Abstract

The invention discloses a visual analysis-based steel-wood combination template welding seam quality detection method, and relates to the related field of machine vision and image processing, the method comprises the following steps: collecting image data of a steel-wood combination template surface welding seam, the images being divided into two types of welding seam defect-free and welding seam defect-having; in view of the reflection difference of different materials of the steel-wood combined template, preprocessing the collected surface weld seam image of the steel-wood combined template, adjusting the brightness distribution of the image surface, and extracting weld seam features; marking welding seam features, training a welding seam quality detection model, and storing model parameters with the best performance as a pre-training model; in a real production environment, inputting a steel-wood combined template surface welding seam image shot in real time into the pre-training model, carrying out deep analysis on welding seam characteristics, and outputting a detection result; and according to a detection result of the model, evaluating the welding seam quality of the steel-wood combined template, and completing a quality detection report. And the welding seam quality detection precision and efficiency are improved.

Description

Technical Field

[0001] The present application relates to the fields related to machine vision and image processing, and in particular to a method for detecting the weld quality of steel-wood composite membrane panels based on visual analysis. Background Art

[0002] With the development of new materials and technologies, new building materials containing two or more components have been widely promoted. Quality testing has become an essential part of new material testing services. As one of the representatives of new building materials, steel-wood composite formwork is particularly important for quality testing. Weld quality testing is a key step in ensuring the safety and reliability of steel-wood composite formwork structures.

[0003] Traditional weld quality inspection methods mainly rely on manual inspection, which evaluates weld quality by visually inspecting the appearance of steel-wood composite formwork. The method is simple but has limited detection accuracy. Traditional inspection methods often use tools and chemical reagents to detect weld defects in steel-wood composite formwork based on chemical and physical principles. For example, penetrants applied to the welding surface are used to observe the color change of the penetrant to determine whether there are weld defects, or the transmission ability of X-rays is used to determine defects in the weld through X-ray images.

[0004] Traditional inspection methods require specialized equipment and personnel, have limited automation, and are inflexible, leading to errors in the identification of weld defects in steel-wood composite formwork. A new approach to improve inspection accuracy is urgently needed. This approach, through visual analysis, can reduce the brightness variations caused by the material properties of steel composite formwork, improve image quality, and extract typical features of weld defects. Deep learning models can then be used to uncover the inherent patterns of these defects, enhancing the stability and robustness of weld quality inspection. Summary of the Invention

[0005] In response to the technical problems of the above-mentioned prior art, this application provides a steel-wood composite membrane panel weld quality detection method based on visual analysis, applies image processing methods to solve the problem of uneven brightness distribution of weld images on the surface of steel-wood composite membrane panels, highlights weld features, introduces deep learning models to deeply analyze weld features, automatically identifies the type of weld defects, and improves the accuracy and efficiency of steel-wood composite membrane panel weld quality detection.

[0006] This application provides a method for detecting the weld quality of steel-wood composite membrane panels based on visual analysis, comprising:

[0007] (1) Collect image data of welds on the surface of steel-wood composite formwork, and divide the images into two categories: "no weld defects" and "with weld defects";

[0008] (2) In view of the difference in reflectivity of different materials of steel-wood composite templates, the collected weld seam images on the steel-wood composite template surface are preprocessed to adjust the brightness distribution of the image surface and extract the weld seam features;

[0009] (3) Label the weld features, train the weld quality detection model, and save the best performing model parameters as the pre-training model;

[0010] (4) In a real production environment, the real-time weld seam images of the steel-wood composite template surface are input into the pre-trained model, the weld seam features are deeply analyzed, and the detection results are output;

[0011] (5) Based on the test results of the model, evaluate the weld quality of the steel-wood composite formwork and complete the quality inspection report.

[0012] Furthermore, the equipment for collecting images of weld seams on the steel-wood composite formwork surface includes an industrial camera, a fixed-focus lens, a ring light source, and a fixing device. The detailed process of collecting images includes:

[0013] Fix the steel-wood combination template on the shooting platform, and install the camera and ring light source in key positions;

[0014] Adjust camera parameters, set camera exposure time and gain value;

[0015] Photograph weld areas with good welding quality and no defects, and photograph various parts of the weld from different angles and positions as "no weld defect" image samples;

[0016] Select areas containing different types of weld defects, adjust camera parameters to highlight defect features, and take multiple images of defective welds as "weld defect image samples";

[0017] The captured images are saved in a computer for subsequent preprocessing.

[0018] Furthermore, the steel-wood composite formwork is made of steel and wood. The different reflective properties of the two materials result in uneven brightness distribution and blurring when analyzing the steel-wood composite formwork image. Therefore, the collected weld seam images of the steel-wood composite formwork surface need to be preprocessed. The preprocessing methods include grayscale conversion, denoising, and contrast enhancement:

[0019] First, the weighted average method is used to convert the collected weld image into a grayscale image. Then, the linear stretching method is used to adjust the grayscale value of the image. The grayscale values ​​of pixels above the preset grayscale upper limit are compressed, and the grayscale values ​​of pixels below the preset grayscale lower limit are expanded, so that the grayscale values ​​of all pixels are remapped to the specified grayscale range.

[0020] Secondly, a denoising method based on partial differential equations is used to reduce noise interference in the weld grayscale image. By constructing a nonlinear diffusion equation, the noise is smoothed during the image diffusion process and the weld boundary is highlighted.

[0021] Finally, the multi-scale Retinex algorithm is applied to the denoised weld grayscale image to decompose the illumination component and reflection component of the image and balance the reflection differences between different materials.

[0022] Furthermore, the enhanced weld grayscale image is segmented using the Otsu threshold method to extract weld features on the surface of the steel-wood composite formwork. This method determines the optimal threshold by maximizing the inter-class variance and segments the weld features. The steps include:

[0023] Count the number of pixels at each gray level in the image to obtain a grayscale histogram;

[0024] Traverse all thresholds T and divide the image into foreground and background. The division is based on the following:

[0025]

[0026] Among them, I(i,j) represents the grayscale value of pixel (i,j);

[0027] Calculate the weights and mean of the foreground and background, where the weight represents the ratio of the number of foreground or background pixels to the total number of pixels;

[0028] Calculate the inter-class variance g between the foreground and background, the calculation formula is:

[0029] g=ω0·ω1·(μ0-μ1) 2

[0030] Among them, ω0 and ω1 represent the weights of the foreground and background respectively; μ0 and μ1 represent the average grayscale values ​​of the foreground and background respectively;

[0031] The threshold that maximizes the inter-class variance is selected as the optimal segmentation threshold, the foreground pixels greater than or equal to the threshold are set to 255, and the background pixels less than the threshold are set to 0, and a binary image is generated as the weld feature.

[0032] Furthermore, the weld quality inspection model adopts the OverLoCK structure, based on the bionic deep stage decomposition strategy, and inspired by the "overview first, then detailed look" mechanism of the human visual system, the network structure is divided into three sub-networks: the backbone network, the overview network, and the focus network.

[0033] The backbone network downsamples the input image through the embedding layer to generate an intermediate-level feature map. This feature map is fed simultaneously into a lightweight overview network and a deeper focus network. The overview network further downsamples the intermediate-level feature map to generate a semantically rich overview feature map that contains a holistic understanding of the image. This feature map is used as a feedback signal and integrated into all modules of the focus network, providing overall contextual prior information. Finally, guided by this contextual prior information, the focus network gradually refines the intermediate-level feature map, expanding the receptive field and obtaining an information-rich high-level feature representation.

[0034] In addition, different types of welds have different sizes and feature information. The convolution kernel with fixed weights is not capable of processing different weld features and its adaptability is limited. Therefore, the dynamic module inside the focusing network adopts context-mixed dynamic convolution to enable the convolution to have the ability to model long-distance dependencies and maintain a strong inductive bias. By calculating the affinity value between each spatial position in the input feature map and the central area of ​​the context prior feature map, a dynamic convolution kernel is generated, and the convolution weights are adaptively adjusted to process different types of weld features.

[0035] Furthermore, the stochastic gradient descent method is used to train the weld quality inspection model. The overview network and the focus network are connected to their respective classifiers to calculate the corresponding classification loss. The classification loss uses the cross entropy loss function.

[0036] During the training process, images with and without weld defects are fed into the model as training samples, enabling the model to determine whether weld defects exist and identify the defect categories.

[0037] The Adam optimizer and cross-validation method are used, and regularization is introduced to reduce the risk of model overfitting;

[0038] When the loss function converges, the model parameters with the best performance are saved as the pre-trained weld quality detection model, which is used to detect the weld quality of the steel-wood composite formwork surface in real time.

[0039] The present invention discloses the following technical effects:

[0040] This paper proposes a visual analysis-based method for inspecting weld seam quality in steel-wood composite formwork. It incorporates multiple image processing techniques to enhance image representation and applies a deep learning model to identify weld defect categories on the surface of steel-wood composite formwork. To highlight the weld areas in the image, the captured weld images are first converted to grayscale. A denoising method based on partial differential equations is then used to remove noise interference. A multi-scale Retinex algorithm is then used to enhance image contrast and balance brightness differences between different materials. Finally, the Otsu algorithm is used to segment the image foreground and background, and the binarized image is used as weld features for input into the weld quality inspection model. This model, inspired by the "overview before detailed" mechanism of the human visual system, consists of a backbone network, an overview network, and a focus network. The backbone network provides contextual prior information to the other two networks. A top-down attention mechanism is employed, along with contextual hybrid dynamic convolution, to extract global information from the weld features, enabling accurate identification of weld defect categories. This paper, based on image processing and deep learning techniques, improves the accuracy and efficiency of weld quality inspection for steel-wood composite formwork. In addition, the present invention provides an efficient detection method for new material detection services, which can be used for surface quality detection of various new materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention are briefly introduced below. Flowcharts are used in this application to illustrate the operations performed according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in precise order. Instead, various steps may be processed in reverse order or simultaneously as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0042] Figure 1 A flow chart of a method for detecting weld quality of steel-wood composite formwork based on visual analysis is provided in an embodiment of the present application.

[0043] Figure 2 This is a schematic diagram of the structure of the weld quality detection model provided in the embodiment of the present application.

[0044] Figure 3 Detailed structural diagram of the internal modules of the weld quality detection model provided in the embodiment of the present application. DETAILED DESCRIPTION

[0045] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0046] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0047] In the following description, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or are inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein are for the purpose of describing the embodiments of this application only.

[0048] Example 1: This application example provides a method for detecting weld quality of steel-wood composite templates based on visual analysis. Figure 1 As shown, the method includes:

[0049] Step S10, collecting image data of the welds on the surface of the steel-wood composite formwork, and dividing the images into two categories: "no weld defects" and "with weld defects".

[0050] In this embodiment, the equipment for collecting images of weld seams on the surface of the steel-wood composite formwork includes an industrial camera, a fixed-focus lens, a ring light source, and a fixing device. The detailed process of collecting images includes:

[0051] Fix the steel-wood combination template on the shooting platform, and install the camera and ring light source in key positions:

[0052] Install a 50mm fixed-focus lens on the camera, mount the camera on a tripod, and adjust the camera position and angle so that the weld area is in the center of the camera's field of view; install a ring light source in front of the camera, and adjust the brightness and angle of the light source to ensure that the weld surface is evenly illuminated.

[0053] Adjust the camera parameters and set the camera exposure time and gain value:

[0054] Photograph weld areas with good welding quality and no defects, and photograph various parts of the weld from different angles and positions as "no weld defect" image samples;

[0055] Select areas containing different types of weld defects, adjust camera parameters to highlight the defect features, and take multiple images of defective welds as "with weld defects" image samples:

[0056] In this embodiment, the types of weld defects include cracks, pores, lack of fusion, and undercuts. Multiple images are taken for each defect type to ensure that sufficient training samples are collected and sample diversity is guaranteed.

[0057] Save the captured images to the computer for subsequent preprocessing:

[0058] The surface weld images of the steel-wood composite formwork were uniformly saved in PNG format, and the images were further preprocessed using a computer program.

[0059] In step S20, considering the difference in reflectivity of different materials of the steel-wood composite formwork, the collected weld seam image on the surface of the steel-wood composite formwork is preprocessed to adjust the brightness distribution of the image surface and extract weld seam features.

[0060] In this embodiment, the preprocessing method of the collected weld seam image on the surface of the steel-wood composite formwork includes grayscale conversion, denoising and contrast enhancement.

[0061] First, the weighted average method is used to convert the collected weld image into a grayscale image. The method is expressed as follows:

[0062] Y=0.299R+0.587G+0.114B

[0063] Where Y is the grayscale value; R, G, and B are the pixel values ​​of the red, green, and blue channels, respectively. A linear stretching method is used to adjust the grayscale value of the image. The grayscale values ​​of pixels above the preset grayscale upper limit are compressed, and the grayscale values ​​of pixels below the preset grayscale lower limit are expanded, so that the grayscale values ​​of all pixels are remapped to the specified grayscale range:

[0064] Find the minimum and maximum grayscale values ​​in the weld grayscale image, set the target grayscale interval to [0, 255], and apply the following formula to transform the grayscale value I(x, y) of each pixel in the image:

[0065]

[0066] Among them, I new (x, y) represents the new grayscale value, min and max represent the minimum grayscale value and the maximum grayscale value respectively;

[0067] Secondly, a denoising method based on partial differential equations is used to reduce noise interference in the weld grayscale image. By constructing a nonlinear diffusion equation, the noise is smoothed during the image diffusion process. The nonlinear diffusion equation retains more weld area information while denoising and highlights the weld boundary.

[0068] In this embodiment, the nonlinear diffusion equation is the Perona-Malik equation, which enhances diffusion in flat areas of the image to remove noise and weakens diffusion in edge areas to preserve edge information. The equation is expressed as:

[0069]

[0070] in, is the rate of change of the image at time t, indicating that the image diffuses over time; u is the grayscale value of the image at the (x, y) position; represents the gradient of the image, and |·| represents the modulo operation; is the nonlinear diffusion coefficient. In this embodiment, the coefficient is calculated as follows:

[0071]

[0072] Among them, K is the threshold parameter used to control the intensity of the gradient response.

[0073] In this embodiment, the detailed process of denoising the weld grayscale image includes:

[0074] Read the noisy weld grayscale image, set the initial time t = 0, select the diffusion coefficient calculation formula and threshold parameters; at each time step, calculate the image gradient and gradient modulus; calculate the diffusion coefficient based on the gradient modulus; update the image according to the nonlinear differential equation:

[0075]

[0076] Where Δt is the time step; u t and u t+1 represent the image grayscale value at time t and the updated image grayscale value respectively; repeat the above steps until the predetermined number of iterations is reached to obtain the denoised weld grayscale image.

[0077] Finally, the multi-scale Retinex algorithm is applied to the denoised weld grayscale image to decompose the image's illumination and reflection components and balance the reflection differences between different materials. The multi-scale Retinex algorithm steps include:

[0078] Use Gaussian filters of different scales to smooth the image and obtain illumination components of different scales. The formula is expressed as:

[0079] L σ =G σ *I

[0080] Where I represents the input image, G σ represents a Gaussian filter with a standard deviation of σ, * represents a convolution operation, and L σrepresents the illumination component of the image at a scale with a standard deviation of σ;

[0081] According to the illumination component of each scale, the reflection component is calculated. The calculation formula of the reflection component is as follows:

[0082] R σ =logI-logL σ

[0083] Among them, the reflection component of the image at the scale with standard deviation σ;

[0084] The reflection components of different scales are weighted and fused to obtain the final enhanced image R MSR :

[0085]

[0086] Where N is the number of scales, σ n is the standard deviation of the nth scale;

[0087] After preprocessing the weld image on the steel-wood composite formwork surface, the enhanced weld grayscale image is segmented using the Otsu threshold method to extract the weld features on the steel-wood composite formwork surface. This method determines the optimal threshold by maximizing the inter-class variance and segments the weld features. The steps include:

[0088] Count the number of pixels at each gray level in the image to obtain a grayscale histogram;

[0089] Traverse all thresholds T and divide the image into foreground and background. The division is based on the following:

[0090]

[0091] Among them, I(i,j) represents the grayscale value of pixel (i,j);

[0092] Calculate the weights and mean of the foreground and background, where the weight represents the ratio of the number of foreground or background pixels to the total number of pixels;

[0093] Calculate the inter-class variance g between the foreground and background, the calculation formula is:

[0094] g=ω0·ω1·(μ0-μ1) 2

[0095] Among them, ω0 and ω1 represent the weights of the foreground and background respectively; μ0 and μ1 represent the average grayscale values ​​of the foreground and background respectively;

[0096] The threshold that maximizes the inter-class variance is selected as the optimal segmentation threshold, the foreground pixels greater than or equal to the threshold are set to 255, and the background pixels less than the threshold are set to 0, and a binary image is generated as the weld feature.

[0097] Step S30: annotate the weld features, train the weld quality detection model, and save the model parameters with the best performance as a pre-trained model.

[0098] In this embodiment, a labeling tool is used to label the weld features, mark whether there are weld defects, and mark the specific weld defect category, and each label value is represented by an integer.

[0099] The weld quality inspection model adopts the OverLoCK structure, which divides the network into three sub-networks: the backbone network, the overview network, and the focus network:

[0100] The backbone network downsamples the input image through the embedding layer to generate a feature map at an intermediate level; this feature map is fed into both a lightweight overview network and a deeper focus network.

[0101] The overview network further downsamples the feature maps of the intermediate layers to generate a semantically rich overview feature map that contains a holistic understanding of the image. At the same time, this feature map is used as a feedback signal and fused into all modules of the focusing network to provide overall contextual prior information.

[0102] Finally, under the guidance of contextual prior information, the focusing network gradually refines the feature maps of the intermediate levels, expands the receptive field, and obtains information-rich high-level feature representations.

[0103] The dynamic module inside the focusing network adopts context-mixed dynamic convolution to make the convolution adaptive and extract information about different types of weld features. The dynamic convolution kernel is generated by calculating the affinity between each position in the input feature map and the central area in the context prior feature map.

[0104] The stochastic gradient descent method is used to train the weld quality inspection model. The overview network and the focus network are connected to their respective classifiers to calculate the corresponding classification loss. The classification loss uses the cross entropy loss function.

[0105] During the training process, images with and without weld defects are fed into the model as training samples, enabling the model to determine whether weld defects exist and identify the defect categories.

[0106] The Adam optimizer and cross-validation method are used, and regularization is introduced to reduce the risk of model overfitting;

[0107] When the loss function converges, the model parameters with the best performance are saved as the pre-trained weld quality detection model, which is used to detect the weld quality of the steel-wood composite formwork surface in real time.

[0108] Step S40: In a real production environment, the real-time image of the weld seam on the surface of the steel-wood composite template is input into the pre-trained model, the weld seam features are deeply analyzed, and the detection results are output.

[0109] Step S50: Evaluate the weld quality of the steel-wood composite formwork based on the test results of the model and complete a quality test report.

[0110] Example 2: This application embodiment provides a method for detecting weld quality of steel-wood composite templates based on visual analysis, which uses a weld quality detection model to identify welds and classify welds. The detailed structure of the model is as follows: Figure 2 As shown in the figure, the change process of the feature map dimension and the number of each module are marked:

[0111] In this embodiment, the weld quality inspection model adopts the OverLoCK structure, which includes three sub-networks: the backbone network, the overview network, and the focus network. The backbone network provides contextual prior information to the overview network and the focus network, guiding the network's feature learning process.

[0112] The backbone network and overview network are composed of embedding layers and basic modules. In the backbone network, the input weld features are downsampled through three embedding layers. After each downsampling, the feature map size is reduced to half of the original size. The downsampled features are sent to the basic module to extract feature information. The detailed structure of this module is shown in the figure. Figure 3 -(a) shows:

[0113] The input features are first input into a residual depthwise separable convolution with a convolution kernel size of 3×3 for local perception. The output features are sequentially passed through layer normalization, dilated re-parameterized convolution layer, SE layer and convolution feed-forward layer. Fine-grained spatial information is extracted by compressing and then expanding the feature channels, and the SE layer increases the attention to features on important channels.

[0114] Among them, the dilated re-parameterized convolution introduces a multi-branch structure in the training stage to enhance the learning ability of convolution. In the inference stage, the parameters in the multi-branch structure are merged into a convolution kernel, and holes are added to the convolution kernel. The expansion coefficient is set to expand the receptive field without increasing the parameters. The re-parameterized convolution ensures performance while significantly reducing the number of parameters of the model.

[0115] The feature map finally output by the backbone network is used as the intermediate-level feature to guide the learning process of the overview network and the focus network. In addition, the output feature map of the overview network is used as contextual prior information to guide the focus network to extract deeper feature information.

[0116] The focusing network adopts a more complex dynamic module, the detailed structure of which is as follows Figure 3 -(b) shows:

[0117] It consists primarily of residual depthwise separable convolutions, gated dynamic spatial aggregators, and convolutional feedforward layers. This module concatenates the output features of the backbone network and the overview network along the channel dimension, introducing contextual prior information. It then applies convolution to the concatenated features to extract feature information. A residual is added after each convolution layer to supplement the original information and stabilize the model's training process.

[0118] The output feature map of the convolutional feedforward layer is divided, one part is used as the contextual prior information of the next dynamic module, and the other part is weighted and added to the input features of the module as the output feature of the module. The weighted weight is a dynamic weight obtained by model adaptive learning.

[0119] The gated dynamic spatial aggregator consists of two parallel branches. The upper branch introduces contextual hybrid dynamic convolution. By calculating the affinity value between each spatial position in the input feature map and the center of the global context feature map region, it generates a dynamic convolution kernel and injects contextual information into the convolution kernel weights. The information at each spatial position in the feature map is modulated by the approximate global information collected by the region center, effectively modeling long-range dependencies. The lower branch includes a 1×1 convolution and a SiLU activation function. The output features of the two branches are multiplied point by point and then output through a 1×1 convolution.

[0120] The high-level feature map output by the focusing network contains rich information about the weld features. This feature is fed into the classifier, the loss function is calculated with the label information, and the parameters of the model are updated through back propagation.

[0121] In this embodiment, the output features of the overview network and the label information are additionally used to calculate the loss function, and multi-scale feature information is introduced to guide the training process of the model.

[0122] The above specific embodiments do not constitute a limitation to the scope of protection of this application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of this application should be included in the scope of protection of this application. In some cases, the actions or steps recorded in this application can be performed in an order different from that in the embodiments and can still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for detecting weld quality of steel-wood composite formwork based on visual analysis, characterized in that: The method comprises: (1) Collect image data of welds on the surface of steel-wood composite formwork, and divide the images into two categories: "no weld defects" and "with weld defects"; (2) In view of the difference in reflectivity of different materials of steel-wood composite templates, the collected weld seam images on the steel-wood composite template surface are preprocessed to adjust the brightness distribution of the image surface and extract the weld seam features; (3) Label the weld features, train the weld quality detection model, and save the best performing model parameters as the pre-training model; (4) In a real production environment, the real-time weld seam images of the steel-wood composite template surface are input into the pre-trained model, the weld seam features are deeply analyzed, and the detection results are output; (5) Based on the test results of the model, evaluate the weld quality of the steel-wood composite formwork and complete the quality inspection report.

2. A method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 1, characterized in that: In step (1), the equipment for collecting the weld seam image on the surface of the steel-wood composite template includes an industrial camera, a fixed-focus lens, a ring light source, and a fixing device; the detailed process of collecting the image includes: Fix the steel-wood combination template on the shooting platform, and install the camera and ring light source in key positions; Adjust camera parameters, set camera exposure time and gain value; Photograph weld areas with good welding quality and no defects, and photograph various parts of the weld from different angles and positions as "no weld defect" image samples; Select areas containing different types of weld defects, adjust camera parameters to highlight the defect features, and capture multiple images of defective welds as "weld defect image samples"; The captured images are saved in a computer for subsequent preprocessing.

3. The method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 1, characterized in that: In step (2), the steps of pre-processing the collected weld seam image on the surface of the steel-wood composite template are as follows: First, the weighted average method is used to convert the collected weld image into a grayscale image. Then, the linear stretching method is used to adjust the grayscale value of the image. The grayscale values ​​of pixels above the preset grayscale upper limit are compressed, and the grayscale values ​​of pixels below the preset grayscale lower limit are expanded, so that the grayscale values ​​of all pixels are remapped to the specified grayscale range. Secondly, a denoising method based on partial differential equations is used to reduce noise interference in the weld grayscale image. By constructing a nonlinear diffusion equation, the noise is smoothed during the image diffusion process and the weld boundary is highlighted. Finally, the multi-scale Retinex algorithm is applied to the denoised weld grayscale image to decompose the illumination component and reflection component of the image and balance the reflection differences between different materials.

4. A method for detecting weld quality of steel-wood composite formwork based on visual analysis as claimed in claim 3, characterized in that: The nonlinear diffusion equation is the Perona-Malik equation expressed as: in, is the rate of change of the image at time t, indicating that the image diffuses over time; u is the grayscale value of the image at the (x, y) position; represents the gradient of the image, and |·| represents the modulo operation; is the nonlinear diffusion coefficient, which is calculated as follows: Among them, K is the threshold parameter used to control the intensity of the gradient response.

5. The method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 3, characterized in that: The detailed process of implementing weld grayscale image denoising based on partial differential equations includes: Read the noisy weld grayscale image, set the initial time t = 0, select the diffusion coefficient calculation formula and threshold parameters; at each time step, calculate the image gradient and gradient modulus; calculate the diffusion coefficient based on the gradient modulus; update the image according to the nonlinear differential equation: Where Δt is the time step; u t and u t+1 represent the grayscale value of the image at time t and the grayscale value of the image at time t+1 after the update respectively; repeat the above steps until the predetermined number of iterations is reached to obtain the denoised weld grayscale image.

6. The method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 3, characterized in that: The multi-scale Retinex algorithm steps include: Use Gaussian filters of different scales to smooth the image and obtain illumination components of different scales. The formula is expressed as: L σ =G σ *I Where I represents the input image, G σ represents a Gaussian filter with a standard deviation of σ, * represents a convolution operation, and L σ represents the illumination component of the image at a scale with a standard deviation of σ; According to the illumination component of each scale, the reflection component is calculated. The calculation formula of the reflection component is as follows: R σ =logI-logL σ Among them, the reflection component of the image at the scale with standard deviation σ; The reflection components of different scales are weighted and fused to obtain the final enhanced image R MSR : Where N is the number of scales, σ n is the standard deviation of the nth scale.

7. The method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 1, characterized in that: In step (2), the enhanced weld grayscale image is segmented using the Otsu threshold method to extract weld features on the surface of the steel-wood composite template. This method determines the optimal threshold by maximizing the inter-class variance and segments the weld features. The steps include: Count the number of pixels at each gray level in the image to obtain a grayscale histogram; Traverse all thresholds T and divide the image into foreground and background. The division is based on the following: Among them, I(i,j) represents the grayscale value of pixel (i,j); Calculate the weights and mean of the foreground and background, where the weight represents the ratio of the number of foreground or background pixels to the total number of pixels; Calculate the inter-class variance g between the foreground and background, the calculation formula is: g=ω0·ω1·(μ0-μ1) 2 Among them, ω0 and ω1 represent the weights of the foreground and background respectively; ω0 and μ1 represent the average grayscale values ​​of the foreground and background respectively; The threshold that maximizes the inter-class variance is selected as the optimal segmentation threshold, the foreground pixels greater than or equal to the threshold are set to 255, and the background pixels less than the threshold are set to 0, and a binary image is generated as the weld feature.

8. The method for detecting weld quality of steel-wood composite formwork based on visual analysis according to claim 1, characterized in that: In step (3), the weld quality inspection model adopts the OverLoCK structure, which includes three sub-networks: backbone network, overview network and focus network: The backbone network downsamples the input image through the embedding layer to generate a feature map at the intermediate level; this feature map is simultaneously fed into a lightweight overview network and a deeper focusing network; the overview network further downsamples the feature map at the intermediate level to generate a semantically rich overview feature map, which is used as a feedback signal and fused into all modules of the focusing network to provide overall contextual prior information; finally, under the guidance of the contextual prior information, the focusing network gradually refines the feature map at the intermediate level, expands the receptive field, and obtains information-rich high-level feature representations.

9. A method for detecting weld quality of steel-wood composite formwork based on visual analysis as claimed in claim 8, characterized in that: The stochastic gradient descent method is used to train the weld quality inspection model. The overview network and the focus network are connected to their respective classifiers to calculate the corresponding classification loss. The classification loss uses the cross entropy loss function. During the training process, images with and without weld defects are fed into the model as training samples, enabling the model to determine the presence of weld defects and identify the defect categories. The Adam optimizer and cross-validation method are used, and regularization is introduced to reduce the risk of model overfitting; When the loss function converges, the model parameters with the best performance are saved as the pre-trained weld quality detection model, which is used to detect the weld quality of the steel-wood composite formwork surface in real time.

Images