Generator circuit breaker tank welding point positioning method based on machine vision

Abstract

The application relates to the technical field of image processing, in particular to a generator circuit breaker box welding point positioning method based on machine vision; the method comprises the following steps: acquiring a to-be-processed image of a generator circuit breaker box, and performing multi-scale decomposition on the to-be-processed image to extract a high-frequency detail layer representing surface texture and welding point morphology; box structure features are extracted from the high-frequency detail layer, matched with a pre-stored box three-dimensional model containing ideal welding point positions, a pose transformation matrix of the three-dimensional model to a to-be-processed image coordinate space is calculated, and the ideal welding point positions and topological relations in the three-dimensional model are projected to an image space by using the transformation matrix to serve as theoretical constraints of candidate welding points; a random conditional field model is constructed to optimize and screen the candidate welding points, and the final welding point positions are determined by minimizing an energy function of the random conditional field model. The application has the effect of improving the positioning precision of the box welding point positions.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a machine vision-based method for locating welding points in generator circuit breaker housings. Background Technology

[0002] The generator circuit breaker enclosure is a core component in a power system, responsible for both shut-off and protection functions. The welding quality of its enclosure directly affects the safety and structural stability of the entire system. Traditional welding quality inspection relies primarily on manual visual inspection, which is inefficient, labor-intensive, and susceptible to subjective influences, making consistency and reliability difficult to guarantee. With the increasing level of industrial automation, the use of machine vision technology for automatic positioning and inspection of weld points has become a key aspect of improving production efficiency and quality control. Early machine vision methods were typically based on traditional image processing techniques, such as edge detection, Hough transform, or template matching algorithms to identify weld points. However, the surface of generator circuit breaker enclosures often contains oil stains, scratches, reflections, and complex textures inherent in the metal material, all of which severely interfere with the image. Traditional algorithms are highly sensitive to changes in lighting, background noise, and irregular weld point shapes, exhibiting poor robustness and making it difficult to achieve stable and reliable positioning results in complex industrial environments.

[0003] To overcome the limitations of traditional algorithms, researchers have begun applying deep learning, particularly convolutional neural networks (CNNs), to weld point detection tasks. Deep learning-based object detection algorithms, such as YOLO and Faster R-CNN, can automatically learn features from data, exhibiting greater adaptability to variations in weld point morphology and texture. While these methods have improved detection accuracy to some extent, several problems remain. Existing detection networks primarily output bounding boxes of the target, which is insufficient for automated welding or inspection applications requiring precise center point coordinates. When images contain defects such as pits or dents that resemble weld points, the algorithm is prone to misclassification (identifying defects as weld points) or missed detections when the weld point image quality is poor. Summary of the Invention

[0004] To improve the accuracy of weld point detection, this application provides a machine vision-based method for locating weld points in generator circuit breaker enclosures.

[0005] This application provides a machine vision-based method for locating welding points in generator circuit breaker enclosures, employing the following technical solution:

[0006] A machine vision-based method for locating weld points in generator circuit breaker enclosures involves acquiring an image of the generator circuit breaker enclosure to be processed and performing multi-scale decomposition to extract a high-frequency detail layer characterizing surface texture and weld point morphology. This high-frequency detail layer is then input into a pre-trained detection and segmentation network to obtain a saliency probability map, center point offset, and weld point segmentation contour. Combining the saliency probability map, center point offset, and segmentation contour, a set of candidate weld points with confidence scores and segmentation contour information is generated. Finally, the enclosure structural features are extracted from the high-frequency detail layer and registered with a pre-stored 3D model of the enclosure containing ideal weld point locations. The process involves calculating the pose transformation matrix from the 3D model to the coordinate space of the image to be processed, and using this transformation matrix to project the ideal welding point positions and topological relationships in the 3D model onto the image space, serving as theoretical constraints for candidate welding points. A stochastic conditional field model is then constructed to optimize and filter candidate welding points. The univariate potential function of the stochastic conditional field model is set based on the confidence level and contour segmentation quality of the candidate welding points, while the bivariate potential function is set based on the similarity between the actual spatial relationship between any two candidate welding points and the topological relationship in the theoretical constraints. Finally, the welding point positions are determined by minimizing the energy function of the stochastic conditional field model.

[0007] By performing multi-scale decomposition on the box-shaped image and extracting high-frequency detail layers, the microstructural features such as edges and weld textures around the welding points are effectively enhanced, providing a richer visual information foundation for the subsequent detection network. A pre-trained detection and segmentation network is introduced, and the saliency probability map and center point offset are used to correct the localization error of the network's predicted center, further improving the geometric accuracy of candidate welding points. Furthermore, the shape information of candidate points is described using the welding point contours provided by the segmentation branch. Combining the confidence scores of candidate points with the segmentation contours, a structured set of candidate welding points is constructed, providing a quantitative basis for subsequent modeling.

[0008] Furthermore, by extracting structural features such as the box edges and corner lines, and performing pose registration with a 3D structural model containing the ideal welding point locations, the theoretical welding point layout is projected into the image space, constructing spatial topological constraints based on the 3D model, thus providing prior structural information for the conditional random field. The constructed random conditional field model uses confidence and contour quality as a univariate potential function to measure the rationality of retaining candidate points individually; a binary potential function is defined by the deviation between the actual relative position and the theoretical topology to quantify the structural consistency among candidate welding points. The optimal solution obtained through the energy function minimization process balances local accuracy and overall structural rationality, significantly reducing the probability of false detection and false negative detection, achieving highly robust welding point localization under complex interference backgrounds, and providing reliable position information for subsequent automatic welding or inspection tasks.

[0009] Optionally, an image of the generator circuit breaker enclosure is acquired and decomposed at multiple scales to extract a high-frequency detail layer characterizing the surface texture and weld point morphology, including:

[0010] The image to be processed is decomposed using bioorthogonal wavelet transform to obtain multiple high-frequency detail components. The high-frequency detail components at different levels are then fused to obtain the high-frequency detail layer.

[0011] By introducing biorthogonal wavelet transform for multi-scale image decomposition, detailed texture information in each scale direction can be systematically extracted, effectively highlighting high-frequency features such as arc edges and protrusions in the welding point area.

[0012] Optionally, by combining the local extrema, center point offset, and segmentation contour of the saliency probability map, a set of candidate welding points with confidence level and segmentation contour information is generated, including:

[0013] In the saliency probability map, points that are maximal in the neighborhood and whose pixel values ​​are greater than a first preset threshold are identified as initial centers; the initial centers are corrected based on the offset corresponding to the initial centers to obtain candidate welding points; the probability value corresponding to the initial centers in the saliency probability map is used as the confidence level of the candidate welding points; the segmentation contours associated with the candidate welding points are extracted as the segmentation contour information of the candidate welding points.

[0014] Local maxima detection in the saliency map ensures that the candidate weld point region has a locally optimal response, and background noise can be effectively filtered by setting a preset threshold. The center point offset is used to correct the offset between the peak value of the heat map and the actual weld center, improving the positioning accuracy. The extraction of the segmented contour not only provides a shape description, but also provides a quantitative basis for the contour quality evaluation in the subsequent conditional random field. This method provides a high-quality set of candidate weld points for the final screening through a refined candidate point generation strategy, effectively avoiding the identification error caused by inaccurate candidate weld points.

[0015] Optionally, structural features include: box edges and corner lines; extract box structural features from the high-frequency detail layer, register them with a pre-stored 3D box model containing ideal weld point locations, and calculate the pose transformation matrix from the 3D model to the coordinate space of the image to be processed, including:

[0016] Feature detection algorithms are used to extract feature points on the high-frequency detail layer to generate corresponding description vectors; the plane where the welding point is located in the 3D model is projected into a 2D view, and the description vectors extracted from the image are compared with the description vectors at the corresponding positions in the 2D view of the model to establish an initial set of matching point pairs; a homography matrix is ​​obtained from the matching point pairs using a random sampling consensus algorithm, which serves as the pose transformation matrix.

[0017] By projecting the ideal 3D weld point into the image space and establishing theoretical position constraints, the candidate point no longer depends solely on the network output, but has a standard answer supported by the structural drawings; thus significantly improving positioning accuracy and structural integrity.

[0018] Optionally, the pre-trained detection and segmentation network adopts an encoder-decoder architecture and integrates a feature pyramid network.

[0019] Optionally, the expression for the univariate potential function of the stochastic conditional field model is:

[0020] In the formula, Indicates candidate welding points The cost of retention; of which The confidence level of the candidate weld point. Candidate weld points for contour segmentation quality The quality of contour segmentation. These are the preset weighting coefficients.

[0021] By organically combining the confidence level of candidate weld points output by the network with the cutting quality of candidate weld points, a more discriminative scoring mechanism is formed.

[0022] Optionally, the expression for the bivariate potential function of the stochastic conditional field model is:

[0023] In the formula, Indicates the retention of candidate welding points and candidate welding points The cost; of which Candidate welding points arrive The actual relative position vector, This is the theoretical relative position vector projected onto the corresponding point in the theoretical constraints. This is a weighting factor used to control the intensity of the penalty.

[0024] Optionally, for any segmentation contour, the product of four times pi and the area of ​​the segmentation contour is used as the first exponent; the square of the perimeter of the segmentation contour is used as the second exponent; and the ratio of the first exponent to the second exponent is used as the contour segmentation quality.

[0025] Optionally, feature detection algorithms such as SIFT can be used to extract structural features of the box from high-frequency detail layer images.

[0026] Optionally, the encoder of the pre-trained detection and segmentation network is ResNet-50.

[0027] This network has a deep number of layers and good feature representation capabilities, which helps to alleviate the gradient vanishing problem in deep network training and enables it to extract richer multi-level semantic information in complex tasks.

[0028] This application has the following technical effects:

[0029] By introducing a 3D model of the box as prior knowledge, the spatial topological relationship of the ideal welding point is projected onto the image to form a theoretical constraint. The random conditional field model is used to globally optimize the detection results, so as to ensure the structural consistency and correctness of the final positioning result. Thus, high-precision and robust positioning of welding points is achieved in complex industrial backgrounds. Attached Figure Description

[0030] Figure 1 This is a flowchart of the machine vision-based generator circuit breaker enclosure welding point positioning method of this application. Detailed Implementation

[0031] Reference Figure 1 This application discloses a machine vision-based method for locating welding points in generator circuit breaker housings, including steps S1-S4:

[0032] S1: Acquire the image of the generator circuit breaker enclosure to be processed, and decompose it at multiple scales to extract the high-frequency detail layer that represents the surface texture and weld point morphology.

[0033] First, images of the circuit breaker enclosure are acquired using devices such as cameras, and then preprocessed to generate the image to be processed. The preprocessing steps are standard techniques in the field, including noise reduction filtering, primarily used to improve image quality. These steps are also standard techniques in the field and will not be elaborated upon here.

[0034] Then, the image to be processed is decomposed by bioorthogonal wavelet transform, which yields a low-frequency approximation component and multiple high-frequency detail components of different scales and directions; the high-frequency detail components include horizontal details, vertical details and diagonal details.

[0035] For example, a resolution of The frontal image of the pixel box is processed using a biorthogonal wavelet basis (bior4.4) with good symmetry and tight support properties. After a three-level decomposition, nine detail sub-band images are obtained, namely the horizontal, vertical, and diagonal components from the first to the third level.

[0036] Finally, the high-frequency detail components from different levels are fused to obtain a high-frequency detail layer. During the fusion process, weighted fusion can be performed based on the high-frequency detail components; that is, the value of each pixel in the high-frequency detail layer is the weighted sum of the pixel values ​​at corresponding positions in multiple sub-band images. The weight coefficients can be set empirically. For example, to highlight the subtle edges of the weld point contour, a higher weight, such as 0.5, can be assigned to the first-level detail component, while the weights for the second and third levels are set to 0.3 and 0.2 respectively. High-frequency information such as the subtle arc-shaped textures around the weld point, the protrusions formed by metal deposition, and surface scratches are significantly enhanced, forming a high-frequency detail layer focused on structural and textural features, providing high-quality input for subsequent feature extraction and model registration.

[0037] S2: Input the high-frequency detail layer into the pre-trained detection and segmentation network to obtain the saliency probability map, center point offset, and segmentation contour of the weld point; combine the saliency probability map, center point offset, and segmentation contour to generate a set of candidate weld points with confidence and segmentation contour information.

[0038] In this embodiment, the detection and segmentation network adopts an encoder-decoder structure and integrates a feature pyramid network; the encoder can be a ResNet-50; the detection branch of the detection and segmentation network is mainly used to generate a saliency probability map in the form of a heatmap, where the value of each pixel in the saliency probability map represents the probability that the point is the center of the welding point. Specifically, since the peak position of the saliency heatmap may have a slight deviation from the true geometric center of the welding point, the network simultaneously outputs an offset map to fine-tune the center position of the welding point.

[0039] During the operation of ResNet-50, when an input image is shown... When processing a box image of pixels, the encoder downsamples layer by layer, generating a series of feature maps at different resolutions, for example, from layer C2. Go to C5 floor The deep feature map C5 contains rich semantic information but has blurred spatial details, while the shallow feature map C2 retains precise location information but has weaker semantics. To enable the network to simultaneously detect both small and large solder joints, the feature pyramid network in the ResNet-50 model plays a crucial role. It upsamples high-level semantic information step by step through a top-down path and fuses it with the next layer's feature map through lateral connections. For example, the upsampled features of C5 are added to the features of C4 to generate a new feature map P4, which possesses both the semantics and localization capabilities of C5. This process continues, generating a set of feature pyramids from P2 to P5. The decoder then performs the final prediction on the pyramid feature maps that fuse multi-scale information. This allows the network to detect small solder joints on high-resolution feature maps and identify large or clustered solder joints on low-resolution feature maps within a unified framework, significantly improving detection accuracy and robustness.

[0040] The segmentation branch is mainly used to output a binary segmentation mask, which accurately identifies the segmentation contour of the welding point;

[0041] In the saliency probability map, points that are maximal values ​​in their neighborhood and whose pixel values ​​are greater than a first preset threshold are identified as initial centers. For example, assuming the saliency probability map is a grayscale image of the same size as the original image, with pixel values ​​ranging from 0 to 1, a first preset threshold of, for example, 0.7 is set to filter out most of the background area. Local maximum points are searched within a sliding window, for example, 5x5. If a pixel has a value of 0.88, which is greater than all other pixel values ​​in its neighborhood and also greater than 0.7, then this point, for example, with coordinates (150, 230), is determined as an initial center. Candidate welding points are obtained by correcting the center point offset at the corresponding position. For example, if the initial center corresponds to a center point offset of (Δx, Δy), the initial center coordinates (x, y) are adjusted to (x + Δx, y + Δy) to obtain the corrected center point, i.e., the candidate welding point. The segmentation contour associated with the candidate welding point is extracted, and the probability value corresponding to the initial center in the saliency probability map is used as the confidence level of the candidate welding point. Thus, each candidate weld point contains three core pieces of information: location coordinates, complete profile, and a quantified confidence level.

[0042] S3: Extract structural features such as box edges and corner lines from the high-frequency detail layer, register them with the pre-stored 3D box model containing the ideal welding point positions, calculate the pose transformation matrix from the 3D model to the coordinate space of the image to be processed, and use the transformation matrix to project the ideal welding point positions and topological relationships in the 3D model to the image space as theoretical constraints for candidate welding points.

[0043] Structural features such as box edges and corner lines are extracted from the high-frequency detail layer; the plane where the welding points are located in the three-dimensional model (the three-dimensional model can be a CAD three-dimensional model) is projected to a two-dimensional space and matched with the feature points corresponding to the structural features in the high-frequency detail layer to obtain matching point pairs; a homography matrix is ​​obtained from the matching point pairs using a random sampling consensus algorithm, and the homography matrix is ​​used as the pose transformation matrix.

[0044] Specifically, the purpose of this step is to precisely align the 3D design blueprint of the enclosure onto the currently captured 2D image, thereby determining the theoretically correct locations of the welding points. On the previously generated high-frequency detail layer image, feature detection algorithms such as ORB or SIFT are used to extract multiple stable feature points, such as structural features like enclosure edges and corner lines, and a descriptive vector describing the appearance of its local neighborhood is generated for each point. The key planes of the enclosure, typically the front panel where the welding points are located, are extracted from the 3D model and projected into an ideal 2D view. By comparing the descriptive vectors extracted from the image with the corresponding descriptive vectors in the 2D view of the model, an initial set of matching point pairs is established; for example, 80 potential matches are found. Due to the existence of false matches, a Random Sampling Consensus (RANSAC) algorithm is used to refine these matches. RANSAC repeatedly selects a small subset, such as 4 pairs of matching points, calculates the homography transformation matrix between them, and uses this matrix to verify the consistency of the remaining matching points. Ultimately, the homography matrix that obtains the most matching point support, such as 65 pairs of inlier support out of 80, is selected as the final pose transformation matrix. This pose transformation matrix can accurately map any point on the 3D model to the pixel coordinates of a higher frequency detail layer image.

[0045] S4: Construct a stochastic conditional field model to optimize and screen candidate welding points. The univariate potential function of the stochastic conditional field model is set according to the confidence level and contour segmentation quality of the candidate welding points, and the bivariate potential function is set according to the similarity between the actual spatial relationship between any two candidate welding points and the topological relationship in the theoretical constraints. The final welding point position is determined by minimizing the energy function of the stochastic conditional field model.

[0046] First, each candidate welding point is considered as a node in the graph model. Each node can take a value of 1 to indicate retention or 0 to indicate rejection. The univariate potential function is related to the cost of marking a node as 1. Its cost is inversely proportional to the confidence of the candidate welding point, and directly proportional to the contour segmentation quality of geometric indicators such as the roundness and area of ​​the segmented contour. That is, the higher the confidence and the better the contour segmentation quality, the lower the cost of retaining the point.

[0047] Then, the binary potential function is determined. The binary potential function is defined between any two nodes and is used to measure the difference between the actual distance and relative direction of the two candidate welding points in the image and the distance and relative direction between the theoretical welding point pair obtained by projection. The smaller the difference, the more they conform to the overall structure, and the lower the cost of keeping this pair of nodes.

[0048] Finally, a graph cut algorithm, such as the maximum flow minimum cut algorithm, is used to solve for the minimum value of the energy function of the stochastic conditional field model (the energy function is the sum of the univariate potential function and the bivariate potential function in the above steps), thereby obtaining a set of welding points with the optimal overall layout, which is the final positioning result.

[0049] Specifically, for each initial candidate welding point Its univariate potential function can be expressed as:

[0050] In the formula, Indicates candidate welding points The cost of retention; of which The confidence level of the candidate weld point. Candidate weld points for contour segmentation quality The quality of contour segmentation. These are preset weighting coefficients; It is a logarithmic function.

[0051] For each candidate weld point, a cost is assigned, reflecting its probability of being a real weld point. The cost consists of two parts, designed to comprehensively consider multiple clues in the network output. The first part is the negative logarithm of the confidence score, i.e., negative... ; The logarithmic function indicates the confidence level given by the network. The higher the value, the lower the cost of retaining the candidate weld point, indicating a greater likelihood that it is a real weld point.

[0052] The second part introduces prior knowledge about geometric shapes. This is a measure of shape compactness, i.e., contour segmentation quality. For any segmented contour, the product of four times pi and the area of ​​the segmented contour is used as the first exponent; the square of the perimeter of the segmented contour is used as the second exponent, and the ratio of the first exponent to the second exponent is the contour segmentation quality. For a perfect circle, the contour segmentation quality is 1. For irregular or elongated shapes, the value is much less than 1. Real weld joints are typically roughly circular or elliptical, exhibiting high compactness. This is achieved by multiplying by a preset weighting coefficient. For example, a value of 0.5, which can be subtracted from the cost, can reward candidate solder joints with regular shapes, thus further reducing their cost. For instance, a candidate solder joint with a confidence level of 0.9 and a compactness of 0.85 may have a very low retention cost, indicating that it is a high-quality candidate solder joint; while another candidate solder joint with the same confidence level but an odd shape and a compactness of only 0.4 will have a much higher univariate potential energy, indicating that it may be an artifact or interference.

[0053] For any two candidate welding points and Its binary potential function can be expressed as:

[0054] In the formula, Indicates the retention of candidate welding points and candidate welding points The cost; of which Candidate welding points arrive The actual relative position vector, This is the theoretical relative position vector projected onto the corresponding point in the theoretical constraints. The weighting factors used to control the intensity of the penalty are preset by those skilled in the art based on experience.

[0055] The binary potential function defines the connection cost between any two nodes in the graphical model. Its core idea is to use the fixed spatial layout of the welding points on the box as a strong constraint to eliminate false detections. In the previous steps, the ideal welding point positions in the 3D model were projected onto the image using the pose transformation matrix, thus obtaining the ideal relative position vector between any two theoretical welding points, i.e. .

[0056] For any two candidate weld points detected in the image and We can calculate the actual relative position vector between their center points. The binary potential function calculates the squared Euclidean distance between two vectors (the actual relative position vector and the theoretical relative position vector), representing the degree of deviation between the observed spatial relationship and the theoretical layout. The deviation value is then multiplied by a weighting factor. For example, it can be set to 2.0 to adjust the strength of the theoretical constraint. If two candidate weld points... and If the actual relative positions of the vectors are almost identical to their theoretical positions, then the magnitude of the vector difference is close to zero, and the binary potential energy is also close to zero, indicating that the combination is highly compatible. Conversely, if their relative positions are significantly different from the theory, the binary potential energy will become very large, resulting in a high penalty, causing the model to tend to abandon this unreasonable combination in the final solution.

[0057] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A machine vision based method for locating welds on a generator circuit breaker tank, the method comprising: The method comprises the following steps: obtaining a to-be-processed image of a generator circuit breaker box, and performing multi-scale decomposition on the to-be-processed image to extract a high-frequency detail layer representing surface texture and weld point morphology; inputting the high-frequency detail layer into a pre-trained detection and segmentation network to obtain a saliency probability map, a center point offset and a weld point segmentation contour; combining the saliency probability map, the center point offset and the segmentation contour to generate a group of candidate weld points with confidence and segmentation contour information; extracting box structure features from the high-frequency detail layer, and registering the box structure features with a pre-stored three-dimensional model of the box containing ideal weld point positions to calculate a pose transformation matrix of the three-dimensional model to a coordinate space of the to-be-processed image, and projecting ideal weld point positions and topological relationships in the three-dimensional model to the image space by using the pose transformation matrix as theoretical constraints of the candidate weld points; constructing a random conditional field model to optimize and screen the candidate weld points, wherein a unary potential function of the random conditional field model is set according to confidence of the candidate weld points and segmentation quality of the candidate weld points, and a binary potential function is set according to similarity between actual spatial relationships between any two candidate weld points and the topological relationships in the theoretical constraints; and determining a final weld point position by minimizing an energy function of the random conditional field model. For any one segmentation contour, a product of four times of a circular constant and an area of the segmentation contour is taken as a first index; a square of a perimeter of the segmentation contour is taken as a second index, and a ratio of the first index to the second index is taken as the segmentation quality of the contour.

2. The machine vision-based generator circuit breaker tank weld point positioning method of claim 1, wherein, obtaining a to-be-processed image of a generator circuit breaker box, and performing multi-scale decomposition on the to-be-processed image to extract a high-frequency detail layer representing surface texture and weld point morphology, comprising: performing decomposition on the to-be-processed image by using biorthogonal wavelet transform to obtain a plurality of high-frequency detail components, and fusing the high-frequency detail components at different levels to obtain the high-frequency detail layer.

3. The machine vision-based generator circuit breaker tank weld point positioning method of claim 1, wherein, combining the saliency probability map, the center point offset and the segmentation contour to generate a group of candidate weld points with confidence and segmentation contour information, comprising: in the saliency probability map, identifying a point as an initial center if the point is a maximum value in a neighborhood and a pixel value of the point is greater than a first preset threshold; correcting the initial center based on an offset corresponding to the initial center to obtain a candidate weld point; taking a probability value corresponding to the initial center in the saliency probability map as confidence of the candidate weld point; and extracting a segmentation contour associated with the candidate weld point as segmentation contour information of the candidate weld point.

4. The machine vision-based generator circuit breaker tank weld point positioning method of claim 1, wherein, The structure features include: edges and corner lines of the box; extracting the box structure features from the high-frequency detail layer, and registering the box structure features with a pre-stored three-dimensional model of the box containing ideal weld point positions to calculate a pose transformation matrix of the three-dimensional model to a coordinate space of the to-be-processed image, comprising: using a feature detection algorithm to extract feature points on the high-frequency detail layer to generate corresponding description vectors; projecting a plane where the weld point is located in the three-dimensional model as a two-dimensional view, comparing the description vectors extracted from the image with the description vectors at corresponding positions in the two-dimensional view of the model to establish a group of initial matching point pairs; and using a random sample consensus algorithm to solve a homography matrix from the matching point pairs as the pose transformation matrix.

5. The machine vision-based generator circuit breaker enclosure weld joint positioning method of claim 1, wherein, The pre-trained detection and segmentation network adopts an encoder-decoder architecture and integrates a feature pyramid network.

6. The machine vision-based generator circuit breaker enclosure weld joint positioning method of claim 1, wherein, The expression of the unary potential function of the random conditional field model is: ; wherein, represents a candidate solder joint reserved cost; wherein is a confidence of the candidate solder joint, is a candidate solder joint is a contour segmentation quality of the segmentation contour of the candidate solder joint, is a preset weight coefficient.

7. The machine vision-based generator circuit breaker enclosure weld joint positioning method of claim 1, wherein, The expression of the binary potential function of the random conditional field model is: ; wherein denotes the cost of retaining a candidate weld point and a candidate weld point ; wherein is the actual relative position vector of the candidate weld point to , is the projected theoretical relative position vector of the corresponding point in the theoretical constraint, is a weight factor for controlling the strength of the penalty.

8. The machine vision-based generator circuit breaker enclosure weld joint positioning method of claim 4, wherein, The structural features of the box are extracted on the high-frequency detail layer image by using a SIFT feature detection algorithm.

9. The machine vision-based generator circuit breaker enclosure weld joint positioning method of claim 5, wherein, The encoder of the pre-trained detection and segmentation network is ResNet-50.

Images