An industrial drawing digitization analysis method based on multi-modal visual fusion and topological search

Abstract

The application discloses an industrial drawing digitization analysis method based on multi-modal visual fusion and topological search, and relates to the technical field of industrial drawing digitization processing. The industrial drawing digitization analysis method based on multi-modal visual fusion and topological search acquires inventory industrial drawing images, performs target detection, outputs element candidate boundary boxes and preliminary categories, performs scale normalization and geometric transformation enhancement on the candidate regions, inputs a trained deep residual network to extract deep semantic features, completes fine-grained arbitration and determines the final category of the element, and sets the pixels in the internal region of the element to zero on the binary layer. The application realizes fine-grained and accurate identification of elements through target detection and deep residual network cascading, constructs a clear search starting point by setting pixels to zero and edge mask, avoids cross misdirection by using motion vector inertia constraint, and crosses pixel breaks and dashed line intervals by using a gap jumping mechanism, and outputs a logically coherent topological relationship of industrial drawings.

Description

Technical Field

[0001] This invention relates to the field of industrial drawing digitization technology, specifically to a method for digitizing and parsing industrial drawings based on multimodal visual fusion and topology search. Background Technology

[0002] In process industries such as petroleum, chemical, and power, process piping and instrumentation diagrams (PBTs) are indispensable core technical documents in the design, construction, operation, maintenance, and renovation of plants. These diagrams use graphic symbols and connecting lines to describe in detail process equipment, instruments, valves, pipelines, and their logical connections. With the deepening of industrial digitalization and digital twin technologies, converting massive amounts of existing paper drawings or scanned documents into structured digital information has become a crucial foundation for industrial upgrading.

[0003] Existing technologies for digitizing industrial drawings typically follow a relatively fixed processing flow. First, high-resolution scanning equipment converts paper drawings into raster images, and algorithms such as Gaussian filtering or median filtering are used for noise reduction preprocessing to minimize noise and artifacts introduced during scanning. Then, a deep learning-based object detection network, such as the YOLO series or SSD single-level detection models, is used to locate and preliminarily classify standard components such as valves, instruments, and pumps across the entire drawing area, outputting the bounding box coordinates and corresponding category labels for each component. For text information extraction, existing solutions usually call a general-purpose optical character recognition engine to perform indiscriminate text scanning across the entire drawing area to obtain text content such as tag numbers and parameter annotations. For reconstructing pipeline topology, existing methods often employ morphological processing techniques based on image thinning or skeleton extraction. This compresses pipeline paths with a certain pixel width into a central skeleton with a single pixel width, and then establishes topological connections between components and pipelines, as well as between pipelines themselves, by traversing the neighborhood connectivity relationships of the skeleton pixels. Finally, a digital report or graph database containing component attributes and connection relationships is generated.

[0004] The limitations of existing technologies include at least the following problems: Due to physical aging and scanning wear, existing industrial drawings generally have pixel-level discontinuities such as broken lines and gaps in dashed lines. Existing methods can only mechanically track along continuous pixels, and are forcibly terminated once a break is reached. This results in a complete pipeline in the process being dissected into multiple unrelated segments in the topology data. Dashed line signal paths are parsed into scattered fragments of line segments, making it difficult to restore them into complete control loops. As a result, the topology structure generated by automated parsing has a large number of breaks and omissions, which require manual redrawing to make it usable. This significantly increases the time and labor costs of the digitization process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for digital parsing of industrial drawings based on multimodal visual fusion and topology search, which solves the problem of topology parsing failure caused by pixel discontinuity defects in existing industrial drawings.

[0006] To achieve the above objectives, this invention provides the following technical solution: a method for digitally analyzing industrial drawings based on multimodal visual fusion and topology search, comprising the following steps: acquiring existing industrial drawing images; performing target detection on the images; outputting candidate bounding boxes containing component spatial location information and preliminary categories; based on the candidate bounding boxes, performing scale normalization and geometric transformation enhancement on the corresponding local image regions; outputting normalized local image regions; inputting the normalized local image regions into a trained deep residual network, which extracts deep semantic features and performs fine-grained semantic re-arbitration to output the final category of the components; based on the candidate bounding boxes, performing analysis on the binarized image layer... The system performs pixel zeroing on the internal region of the element indicated by the candidate bounding box and constructs a probe mask at the edge of the element to determine the search starting point. Starting from the search starting point, a topology search with motion vector inertia constraint and gap jump mechanism is performed within a preset search neighborhood. The motion vector inertia constraint makes the search process prioritize maintaining the same motion direction as the displacement vector at the previous moment to avoid erroneous turning at pipeline intersections. The gap jump mechanism makes the search process perform cross-breakpoint logical reconnection of pixel discontinuities along the current motion vector direction within a preset probe radius until it touches the probe mask of another element. Based on the starting element and target element touched during the topology search, the topology connection relationship between the two elements is generated.

[0007] Furthermore, the specific steps for target detection in the image are as follows: resample the existing industrial drawing images at high resolution and map them to the preset detection space; extract multi-level semantic features using a neural network based on an anchorless architecture; perform coordinate regression and preliminary class probability verification on the industrial components through a decoupled head structure; and output candidate bounding boxes containing the spatial location information of the components and preliminary categories.

[0008] Furthermore, the specific steps for scaling and geometric transformation enhancement of the local image region corresponding to the candidate bounding box are as follows: map the local image region corresponding to the candidate bounding box to a tensor space with a preset standardized pixel size; perform a fixed-angle rotation transformation on the scale-scaled local image region; perform a fixed-position alignment transformation on the scale-scaled local image region; and complete the scale scaling and geometric transformation enhancement processing of the local image region.

[0009] Furthermore, the specific steps for performing fine-grained semantic re-arbitration using the trained deep residual network are as follows: input the local region of the image after scale normalization and geometric transformation enhancement into the trained deep residual network; extract the deep semantic features of the local region of the image by the deep residual network; calculate the probability distribution of the element category based on the deep semantic features; and output the element category with the highest confidence in the probability distribution as the final category of the element.

[0010] Furthermore, the specific steps for constructing a probe mask at the component edge to determine the search starting point are as follows: After setting the pixels in the component's internal region to zero based on the coordinate range of the candidate bounding box, the component edge is processed using a geometric morphology operator; a probe mask with a preset pixel width is constructed at the component edge; a bitwise AND operation is performed on the probe mask and the pipeline pixels in the binarized layer to lock the precise intersection point between the pipeline and the component edge; the precise intersection point is determined as the search starting point for the topology search.

[0011] Furthermore, the specific steps of the motion vector inertial constraint to prioritize maintaining the same motion direction as the displacement vector at the previous moment to avoid erroneous turns at pipeline intersections are as follows: the motion range of the topology search is limited to a preset search neighborhood for iterative detection; in the initial search phase, the component edge normal direction is used as the initial displacement vector, and the subsequent detection step size obtains the displacement vector at the previous moment; the degree of same-direction matching between each candidate detection direction and the displacement vector at the previous moment is calculated, and the candidate detection direction with the highest degree of same-direction matching is assigned the highest search weight; when there is a valid pipeline pixel in the candidate detection direction corresponding to the highest search weight, the next long detection is performed along that direction.

[0012] Furthermore, the matching degree in the same direction is obtained by calculating the dot product of the candidate detection direction vector and the displacement vector at the previous moment. The vector with the largest dot product corresponds to the highest matching degree in the same direction. The specific steps are as follows: obtain the displacement vector at the previous moment and represent it as a two-dimensional direction vector; obtain the candidate detection direction vectors corresponding to each candidate direction within the preset search neighborhood of the current search position and represent each candidate detection direction vector as a two-dimensional direction vector; calculate the dot product of each candidate detection direction vector and the displacement vector at the previous moment; determine the candidate detection direction vector corresponding to the largest dot product as the candidate detection direction vector with the highest matching degree in the same direction.

[0013] Furthermore, the gap skipping mechanism enables the search process to perform cross-breakpoint logical reconnection of pixel discontinuities within a preset detection radius along the current motion vector direction. The specific steps are as follows: When the topology search does not detect a valid pipeline pixel within the preset search neighborhood, it is determined that there is a pixel discontinuity defect at the current search position, and gap skipping detection is initiated; pixel value detection is performed sequentially at each detection step position within the preset detection radius along the current motion vector direction; if a non-zero valid pipeline pixel is detected within the preset detection radius, the path between the current search position and the location of the valid pipeline pixel is completed into a continuous path through logical interpolation; the search position is jumped to the location where the valid pipeline pixel was detected, and the topology search continues.

[0014] Furthermore, the specific steps for sequentially performing pixel value detection at each detection step position within the preset detection radius along the current motion vector direction are as follows: Starting from the current search position, move pixel by pixel along the current motion vector direction, with the number of steps not exceeding the number of pixels corresponding to the preset detection radius; after each step, obtain the pixel value at the current position and determine whether the pixel value is a non-zero valid pipeline pixel; if it is determined to be a valid pipeline pixel, stop moving and record the current position coordinates; if the number of steps reaches the number of pixels corresponding to the preset detection radius and no valid pipeline pixel is detected, mark the current path as a segment to be backtracked and switch the candidate direction to continue detection.

[0015] Furthermore, the specific steps for generating the topological connection relationship between two elements based on the starting element and the target element touched during the topological search process are as follows: When the probe mask corresponding to the search starting point touches the probe mask of another element during the topological search process, the unique identifier of the starting element and the unique identifier of the target element are obtained through the mapping relationship between pixel coordinates and element identifiers; the unique identifier of the starting element and the unique identifier of the target element are recorded as a set of topological connection pairs; the search path length between the search starting point and the touch position is calculated, and the search path length is stored as the path length attribute of the topological connection pair.

[0016] The present invention has the following beneficial effects:

[0017] (1) The industrial drawing digitization analysis method based on multimodal visual fusion and topology search effectively solves the problem of easy confusion of similar components in industrial drawings through the cascaded cooperation of target detection and fine-grained semantic re-arbitration. In the target detection stage, the anchorless neural network quickly scans the entire drawing, locates all potential component areas and makes a preliminary judgment to ensure that no component is missed. After the candidate area is cropped, it is scaled and calibrated, and then sent to the deep residual network for secondary recognition. This network can capture the subtle texture features that are easily lost, such as valve handwheels and instrument scales, and accurately distinguish similar components such as gate valves and check valves with different flow directions. It effectively suppresses the misjudgment of single-level detection, ensures that the component category is consistent with the drawing annotation, and reduces the subsequent asset statistical deviation and manual verification burden.

[0018] (2) The industrial drawing digitization analysis method based on multimodal visual fusion and topology search introduces motion vector inertial constraints in the topology search stage to solve the path ambiguity problem at the intersection of process pipelines. Traditional search is prone to wrong turns at intersections due to lack of directional judgment, and the main pipeline is mistakenly connected to the bypass. However, the present invention refers to the direction of movement at each step of the detection and explores along this direction first, simulating the physical law of straight pipeline laying. It can smoothly cross the intersection area, avoid interference from vertical pipelines, ensure the continuity of the topology tracking direction, and ensure that the analysis connection relationship is consistent with the design of the drawing. Figure 1 This is to prevent logical errors caused by incorrect path reversals.

[0019] (3) The industrial drawing digitization and analysis method based on multimodal visual fusion and topology search overcomes the obstacle to topological connectivity caused by pixel breaks due to drawing quality degradation and dashed line type through the gap jump mechanism. The scanning loss, creases and dashed line intervals of existing drawings can easily cause conventional tracking to terminate due to breakpoints, thus severing the process loop. When the present invention encounters pixel discontinuity, it scans within a preset radius along the current movement direction. After capturing effective pixels, it fills in the breakpoint through logical interpolation, which can completely restore the broken main pipeline and dashed line signal loop, improve the integrity of the topology structure, and significantly reduce the dependence on manual redrawing.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the above advantages at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for digitally analyzing industrial drawings based on multimodal visual fusion and topology search, according to the present invention.

[0022] Figure 2 This is a schematic diagram of coarse detection and fine calibration in the industrial drawing digitization analysis method based on multimodal visual fusion and topology search of the present invention.

[0023] Figure 3This is a schematic diagram of component context search and associated line marking in an industrial drawing digitization and parsing method based on multimodal visual fusion and topology search according to the present invention.

[0024] Figure 4 This is a schematic diagram of the intelligent topology reconstruction algorithm in the industrial drawing digitization and analysis method based on multimodal visual fusion and topology search of the present invention.

[0025] Figure 5 This is a schematic diagram of deep gap jumps and structured engineering drawings in an industrial drawing digitization and analysis method based on multimodal visual fusion and topology search according to the present invention. Detailed Implementation

[0026] Please see Figure 1-5 This invention provides a technical solution: a method for digital analysis of industrial drawings based on multimodal visual fusion and topology search, comprising the following steps: acquiring existing industrial drawing images; performing target detection on the images; outputting candidate bounding boxes containing component spatial location information and preliminary categories; based on the candidate bounding boxes, performing scale normalization and geometric transformation enhancement on the local image regions corresponding to the candidate bounding boxes, outputting normalized local image regions; inputting the normalized local image regions into a trained deep residual network, which extracts deep semantic features and performs fine-grained semantic re-arbitration, outputting the final category of the components; based on the candidate bounding boxes, performing analysis on the binary layer of the image... The pixel inside the element indicated by the bounding box is zeroed out, and a probe mask is built at the edge of the element to determine the search starting point. Starting from the search starting point, a topology search with motion vector inertia constraint and gap jump mechanism is performed in the preset search neighborhood. The motion vector inertia constraint makes the search process prioritize maintaining the same motion direction as the displacement vector at the previous moment to avoid erroneous turning at pipeline intersections. The gap jump mechanism makes the search process perform cross-breakpoint logical reconnection of pixel discontinuities along the current motion vector direction within the preset detection radius until it touches the probe mask of another element. Based on the starting element and target element touched during the topology search, the topology connection relationship between the two elements is generated.

[0027] Among them, the existing industrial drawing images include scanned raster images, which undergo preprocessing after acquisition, specifically:

[0028] Gaussian denoising was performed on existing industrial drawing images to remove scanning noise and JPEG compression artifacts. The Gaussian denoising used a standard Gaussian convolution kernel, and its formula is as follows: ;

[0029] in, The standard deviation is Gaussian, and it is adaptively adjusted according to the image noise intensity.

[0030] After preprocessing, the image is binarized to obtain a binarized layer. The binarization uses an adaptive thresholding method, and the core thresholding formula is: ;

[0031] in, This represents the local grayscale mean. For local grayscale variance, To adjust the coefficients, after binarization, pipeline pixels have non-zero values, while background pixels have zero values;

[0032] The cascaded visual feature enhancement is specifically a cascaded fusion of the target detection branch and the fine-grained re-arbitration branch. The target detection branch outputs a preliminary category to narrow down the fine-grained recognition range, and the fine-grained re-arbitration branch corrects the preliminary category based on deep semantic features.

[0033] During the topology search, the coordinate information of the search path is recorded in real time. ( This is used for calculating the path length of subsequent topology connections, and also marks pipeline intersections encountered during the search process to avoid repeated searches;

[0034] The unique identifier for a component is assigned in the format of "category prefix + probe number", denoted as The category prefix is ​​set according to the final category of the component (e.g., resistor prefix "R", capacitor prefix "C", inductor prefix "L", diode prefix "D"). The detection sequence number increases in the order of target detection to ensure that the identifier of each component is unique. After fine-grained semantic re-arbitration, this identifier is bound to the candidate bounding box and the detection mask to generate a mapping table of pixel coordinates and component identifiers. This is used for generating subsequent topological connection relationships;

[0035] In addition to touching the probe mask of another element, the termination conditions for topology search also include the search position exceeding the range of the binarized layer, the detection of a marked backtracking segment and the absence of new candidate directions to switch. If the latter two termination conditions are triggered, the search path is determined to be invalid, no topology connection relationship is generated, and the path is marked as an invalid path to avoid redundant storage.

[0036] The selection of the preset search neighborhood can be adaptively adjusted according to the pipeline density of the industrial drawings. When the pipeline density is high, a 4-neighborhood is used to reduce false detections at intersections.

[0037] When pipeline density is low, an 8-neighborhood is used to expand the detection range and improve search efficiency; the preset detection radius of the gap skipping mechanism is denoted as... Based on the common pixel discontinuity defect size settings in industrial drawings, the value is usually set to 5-15 pixels. It can be adaptively adjusted according to the scanning quality and pipeline style (such as solid lines and dashed lines) of the actual drawings. The detection radius of dashed pipelines can be appropriately increased to ensure the integrity of cross-interval reconnection.

[0038] During the pre-training process of deep residual networks, the industrial component annotation dataset needs to contain images of industrial drawing components with different scanning qualities and drawing styles, covering common component sizes and orientation deviations to ensure the generalization ability of the pre-trained model. During fine-tuning, mini-batch gradient descent is used, with a batch size of 32 and an initial learning rate of [missing information]. This further optimizes the model's accuracy in fine-grained identification of components in industrial drawings.

[0039] Specifically, the steps for target detection in an image are as follows:

[0040] The existing industrial drawing images are resampled at high resolution and mapped to a preset detection space, specifically as follows:

[0041] Bilinear interpolation algorithm is used to resample existing industrial drawing images to eliminate jagged distortion during image scaling, so that the resolution of the resampled image is adapted to the input requirements of the anchorless neural network, ensuring that the component symbols are clearly distinguishable.

[0042] A neural network based on an anchorless architecture is used to extract multi-level semantic features, specifically as follows:

[0043] The anchorless architecture neural network adopts the YOLO series anchorless model. The network includes a feature extraction backbone network, a neck feature fusion module, and a head detection module. The feature extraction backbone network adopts an alternating structure of convolutional layers and pooling layers to extract multi-level semantic features from the image from shallow to deep layers. Shallow features are used to capture detailed information such as the edges and contours of the components, while deep features are used to capture the overall shape and category features of the components. The neck feature fusion module adopts a feature pyramid structure to fuse semantic features at different levels and output a fused multi-scale feature map.

[0044] The coordinate regression and preliminary class probability test of industrial components are performed using a decoupled head structure, specifically as follows:

[0045] The decoupled header structure is divided into a coordinate regression branch and a category classification branch. The two branches are independent of each other to avoid mutual interference between coordinate regression and category classification. The coordinate regression branch outputs the coordinate parameters of the candidate bounding boxes, including the top left corner of the bounding box. and bottom right corner Coordinates ensure that the bounding box accurately selects the component area;

[0046] Category classification branches use the softmax function The initial category probability of the output element is selected, and the category with the highest probability is chosen as the initial category.

[0047] For example, if an industrial drawing contains three types of components: resistors, capacitors, and inductors, and the probability value corresponding to resistors is the highest, then the initial category is determined to be resistors.

[0048] The output includes candidate bounding boxes and preliminary categories containing component spatial location information, specifically:

[0049] Non-maximum suppression is performed on the candidate bounding boxes output by the coordinate regression branch to remove redundant bounding boxes with overlap exceeding a preset threshold. The overlap is determined using the intersection-overall reunification ratio (IoU), as shown in the formula: ;

[0050] in, , Given two bounding boxes, the bounding box with the highest probability value is retained.

[0051] After processing, the coordinates of each candidate bounding box and its corresponding preliminary category are output, and a unique temporary identifier is assigned to each candidate bounding box. .

[0052] In this implementation scheme, the target detection process of the image first uses high-resolution resampling and bilinear interpolation to effectively avoid jagged distortion that may be introduced by image scaling, so that the outline of the component symbol remains clear and complete. Utilizing the multi-level feature extraction capability of the anchorless neural network, shallow features are responsible for capturing details such as the edge contour of the component, while deep features focus on the overall shape and category representation of the component. The feature pyramid structure is used to fuse information at different scales, thus taking into account both the sensitivity to small symbols and the grasp of the overall structure. The decoupling head separates the coordinate regression and category classification to prevent the two tasks from interfering with each other when updating parameters, making the bounding box localization more accurate and the initial category judgment more focused. Non-maximum suppression is used to remove overlapping and redundant candidate boxes, retaining the results with the highest confidence, and assigning a temporary label to each candidate region, providing an orderly and accurate object to be processed for the matching in the fine-grained re-arbitration stage.

[0053] Specifically, the steps for scale normalization and geometric transformation enhancement of the local image region corresponding to the candidate bounding box are as follows:

[0054] The local image region corresponding to the candidate bounding box is mapped to a tensor space of a preset normalized pixel size, specifically as follows:

[0055] Based on the coordinate information of the candidate bounding box, the corresponding component symbol area is cropped out. A fixed-size scaling algorithm is used to uniformly map the cropped local area to a tensor space with a preset standardized pixel size. During the scaling process, the aspect ratio of the component symbol remains unchanged to avoid symbol deformation. After mapping, the tensor is normalized to normalize the pixel values ​​to the 0-1 range.

[0056] A fixed-angle rotation transformation is performed on a local region of the scale-normalized image, specifically as follows:

[0057] Based on the conventional orientation of components in industrial drawings, fixed rotation angles of 0°, 90°, 180°, and 270° are set. Rotation transformations are then performed on local areas after scale normalization to eliminate the influence of component symbol tilt on recognition. Affine transformations are used to ensure that the component symbol is completely located within the image area after rotation. For example, if the capacitor symbol in the industrial drawing is tilted at 90°, a 90° rotation transformation is performed to rotate it to a horizontal orientation for easier recognition.

[0058] A fixed-position alignment transformation is performed on a local region of the scale-normalized image, specifically as follows:

[0059] An image alignment algorithm is used to align the local regions after rotation transformation to a fixed position, so that the center of the component symbol coincides with the center of the normalized tensor space. At the same time, it is ensured that the key features of the component are located in a preset position. First, the center coordinates of the component symbol are calculated, and then it is translated to the center of the tensor space through translation transformation. For example, when the center of the resistor symbol deviates from the center of the tensor space, the two are made to coincide through translation.

[0060] The process of scaling and geometric transformation to enhance local regions of an image is as follows:

[0061] After scaling, rotating, and aligning the image local area, pixel value verification is performed to remove abnormal pixels generated during the transformation process, ensuring that the local area meets the input requirements of the deep residual network. After verification, the scaled local area is output.

[0062] In this implementation, the candidate region is scaled and geometrically enhanced. First, the cropped component symbols are uniformly scaled to a preset size while maintaining the original aspect ratio to avoid stretching and deformation that could damage the inherent geometric features of the symbols. At the same time, pixel values ​​are normalized to make the data distribution more consistent, providing a stable input basis for the deep network. Fixed-angle rotation correction is performed based on common orientation deviations of components in industrial drawings. For example, tilted capacitor symbols are corrected to a horizontal position, and affine transformation is used to ensure that the main body of the symbol is still within the field of view after rotation, effectively eliminating the interference of orientation differences on subsequent feature extraction. Then, the center of gravity of the symbol is calculated and translated to the center of the image, aligning the key discrimination region with the center of the network's receptive field. This allows the deep residual network to more focus on capturing subtle texture differences located in a fixed region. Pixel verification filters out outliers that may be generated during the transformation process, ensuring the purity of the input data.

[0063] Specifically, the steps for performing fine-grained semantic re-arbitration using the trained deep residual network are as follows:

[0064] The local regions of the image, after scale normalization and geometric transformation enhancement, are input into the trained deep residual network, specifically as follows:

[0065] The deep residual network adopts the ResNet series of networks, which includes an input layer, residual blocks, global average pooling layers and fully connected layers. The trained network has been pre-trained and fine-tuned using an industrial component dataset.

[0066] The pre-training steps are as follows:

[0067] Construct an industrial component annotation dataset containing annotated images of various industrial components, with no less than 1000 annotated images for each component, and divide it into training set, validation set and test set in a 7:2:1 ratio;

[0068] Load pre-trained weights based on the ImageNet dataset, freeze the first 10 layers of the network, and train only the last 5 layers and the fully connected layers;

[0069] The Adam adaptive gradient descent algorithm is used to optimize the parameters, with the initial learning rate set to... The learning rate is reduced to 0.1 times its original value every 10 training epochs, for a total of 50 training epochs, using the cross-entropy loss function. ;

[0070] in, For label values, To predict probabilities, the accuracy of the validation set is monitored in real time. If the accuracy does not improve after 5 consecutive rounds, training is stopped and the optimal model weights are saved.

[0071] The regularized local region is converted into the input format required by the network and then input into the network through the input layer.

[0072] The deep semantic features of local regions in an image are extracted using a deep residual network, specifically as follows:

[0073] Deep residual networks extract features through multiple residual blocks. Each residual block contains two convolutional layers and a shortcut connection to alleviate the gradient vanishing problem. Shallow residual blocks extract fine-grained detail features such as component edges and textures, while deep residual blocks extract global morphology and category differentiation features of components. The extracted deep semantic features are one-dimensional feature vectors, and the feature vectors of different components are significantly different.

[0074] For example, in the characteristic vector of a resistor, the elements corresponding to "parallel pins" and "horizontal lines in the middle" have higher values, while the elements corresponding to "symmetrical poles" have higher values ​​for a capacitor.

[0075] The probability distribution of element categories is calculated based on deep semantic features, specifically as follows:

[0076] The deep semantic feature vector is input into the fully connected layer, mapped to the component category space, and outputs a category score vector. This vector is then converted into a category probability distribution using the softmax function, and the category with the highest probability is selected as the fine-grained recognition result.

[0077] At the same time, a confidence threshold is introduced. When the highest probability is lower than the threshold, it is judged as ambiguous recognition. The preliminary category and fine-grained recognition results are retained for subsequent conflict arbitration.

[0078] The component category with the highest confidence level in the probability distribution is output as the final component category, specifically as follows:

[0079] The fine-grained identification results are compared with the preliminary category of the target detection; if they match, the category is directly used as the final category.

[0080] If there is a discrepancy, conflict arbitration will be initiated, and the category with the higher probability will be selected as the final category.

[0081] If the difference between the two probabilities is less than a preset threshold, it is marked as "awaiting manual confirmation";

[0082] During output, the final category is bound to a temporary identifier of the candidate bounding box to ensure that each bounding box corresponds to a unique final category.

[0083] In this implementation scheme, the trained deep residual network has been pre-trained and fine-tuned using a large-scale industrial component dataset, possessing a stable ability to discriminate fine-grained features of various symbols. During the inference phase, the network effectively overcomes the deep gradient decay problem by utilizing the shortcut connections of residual blocks, enabling shallow structures to capture local details such as edge textures, while deep structures focus on the overall morphological differences of components, thus retaining key clues for distinguishing similar symbols in the feature vector. The fully connected layer maps deep semantics to a category probability distribution and introduces a confidence threshold and conflict arbitration mechanism. When the fine-grained recognition result differs from the initial detected category, a reasonable decision can be made based on the probability. In cases where a decision is difficult to make, the network is marked for manual review, avoiding the erroneous propagation caused by blindly adopting low-confidence results. The confirmed category is bound to the temporary identifier of the candidate bounding box in the output, ensuring that each component has an accurate semantic identity in subsequent topological associations.

[0084] Specifically, the steps for constructing a probe mask at the edge of the component to determine the search starting point are as follows:

[0085] After setting the pixels inside the element to zero based on the coordinate range of the candidate bounding box, the element edges are processed using geometric morphology operators, specifically:

[0086] The internal region of the element is determined based on the coordinates of the candidate bounding box, and all pixel values ​​in the internal region are set to 0 to avoid interfering with the topology search.

[0087] The component edges are processed using expansion and erosion operators. Expansion fills small gaps at the edges, and erosion removes burrs from the edges. Both use rectangular structural elements, resulting in smooth and complete component edges after processing.

[0088] A probe mask with a preset pixel width is constructed at the edge of the component, specifically as follows:

[0089] Based on the morphologically processed edge of the component, a preset pixel width is extended outward to form a closed probe mask. The pixel value of the mask is set to a non-zero value (distinguished from the pipeline pixels) to accurately capture the intersection of the pipeline and the edge of the component.

[0090] For example, the preset width is 3 pixels, which means it extends 1.5 pixels outward from the edge to form a closed mask;

[0091] Perform a bitwise AND operation on the pipeline pixels in the probe mask and binarized layer to pinpoint the precise intersection point between the pipeline and the component edge. Specifically:

[0092] The bitwise AND operation logic is as follows: the pixel value of the probe mask is multiplied by the pixel value of the pipeline. When the result is non-zero, the pixel is the intersection point.

[0093] Perform connected component analysis on the calculation results, remove isolated noise pixels, each connected component corresponds to an intersection point, and calculate the center coordinates of the connected components as the precise intersection point coordinates;

[0094] The exact intersection point is determined as the starting point for the topological search, specifically as follows:

[0095] Filter the locked exact intersections and remove intersections that are too close to each other to avoid duplicate searches;

[0096] Each selected intersection point serves as a search starting point, is assigned a unique search identifier, and its coordinates and corresponding starting element identifier are recorded for subsequent topology connection generation.

[0097] In this implementation, by setting the pixels inside the component to zero, the interference of the component's own lines on the tracking process is physically eliminated. This allows the search operator to focus only on the intersection of the component's edge and the pipeline, avoiding tracking errors or path disorder caused by the internal texture of the component. Morphological processing using dilation and erosion is used to repair and smooth the edges, filling in small gaps commonly found in scanned images and removing burrs, providing a baseline contour for mask construction. Based on this, a probe mask of a preset width is extended outward, and bitwise AND operations are performed with the pipeline pixels in the binarized layer. This can accurately lock the precise intersection coordinates of the pipeline and the component edge. Isolated noise points are filtered out through connected component analysis to ensure that each intersection point is real and reliable. The intersection points are deduplicated and assigned independent search labels, so that the tracking of each pipeline path starts from a clear physical connection point.

[0098] Specifically, the motion vector inertia constraint ensures that the search process prioritizes maintaining the same direction of motion as the displacement vector from the previous moment to avoid erroneous turns at pipeline intersections. The specific steps are as follows:

[0099] The topology search is limited to an iterative exploration within a preset search neighborhood, specifically as follows:

[0100] The preset search neighborhood prioritizes 4 neighborhoods (up, down, left, and right directions) to reduce search redundancy. The iterative detection step size is 1 pixel. After each detection step, the search position is updated and the current position coordinates are recorded.

[0101] In the initial search phase, the displacement vector is taken as the normal direction of the component edge. Subsequent probe steps obtain the displacement vector from the previous moment, specifically as follows:

[0102] In the initial search phase, there is no displacement vector from the previous moment. The normal direction of the component edge is used as the initial displacement vector to ensure that the initial search direction is towards the pipeline extension direction.

[0103] The subsequent detection step size is determined by the displacement vector of the previous two steps, which is the coordinate difference between the position of the previous step and the position of the previous two steps.

[0104] Calculate the degree of same-direction matching between each candidate detection direction and the displacement vector at the previous time step, and assign the candidate detection direction with the highest degree of same-direction matching the highest search weight, specifically as follows:

[0105] The candidate detection directions are all directions within a preset search neighborhood. Combined with the motion vector inertial mechanism, the set of candidate detection directions is determined by the following formula: ;

[0106] in, The displacement vector of the motion at the previous moment. It is a set of 4 neighborhood directions. For candidate detection direction vectors, exclude those that are... opposite direction To avoid accidental detours at pipeline intersections;

[0107] By calculating candidate direction vectors displacement vector from the previous moment The dot product of vectors determines the degree of matching in the same direction. The larger the dot product value, the higher the degree of matching in the same direction. The dot product value is used as the search weight, and the direction with the highest dot product value has the highest weight.

[0108] When there are valid pipeline pixels in the candidate probe direction corresponding to the highest search weight, the next long probe is performed along that direction, specifically as follows:

[0109] Detect whether the pixel in the highest weight direction is a valid pipeline pixel (a non-zero pixel in the binarized layer). If so, probe along that direction and update the search position.

[0110] Otherwise, if other directions are detected according to their weights from high to low, and no valid pixels are found, the gap skipping mechanism is activated.

[0111] The degree of matching in the same direction is obtained by calculating the dot product of the candidate detection direction vector and the displacement vector at the previous moment. The vector dot product with the largest value corresponds to the highest degree of matching in the same direction. The specific steps are as follows:

[0112] Obtain the displacement vector from the previous moment and represent it as a two-dimensional direction vector, specifically as follows:

[0113] The displacement vector at the previous moment is determined by the search position in the previous two steps. It is normalized to a unit direction vector, retaining only the direction information. The normalized vector magnitude is 1.

[0114] For example, the difference between the position coordinates in the previous two steps is (1, 0), and after normalization, it is still (1, 0).

[0115] Obtain the candidate detection direction vectors corresponding to each candidate direction within the preset search neighborhood of the current search position, and represent each candidate detection direction vector as a two-dimensional direction vector, specifically as follows:

[0116] When the preset search neighborhood is 4 neighborhoods, the candidate directions are up, down, left and right, corresponding to four unit direction vectors;

[0117] If it is an 8-neighborhood, four new unit vectors in the diagonal directions are added to cover more directions of pipeline extension;

[0118] Calculate the dot product of each candidate detection direction vector and the displacement vector from the previous moment, specifically as follows:

[0119] Using the two-dimensional vector dot product formula calculate;

[0120] in, The dot product is the angle between two vectors. The value of the dot product ranges from -1 to 1. The closer the dot product value is to 1, the more consistent the directions of the two vectors are, and the higher the degree of matching in the same direction.

[0121] The candidate probe direction vector corresponding to the vector with the largest dot product is determined as the candidate probe direction vector with the highest matching degree in the same direction. Specifically:

[0122] Sort the dot product values ​​of all candidate directions and select the direction with the largest dot product value, which is the direction with the highest degree of matching in the same direction;

[0123] If multiple maximum dot product values ​​exist, the direction that is completely consistent with the displacement vector at the previous moment should be selected first.

[0124] In this implementation, the motion vector inertial constraint, by referencing the historical displacement direction at each detection step, endows the topology search with a path selection preference that conforms to the straight-line laying pattern of the process pipeline. At locations where directional ambiguity is easily generated, such as pipeline intersections, traditional search methods often randomly turn due to a lack of judgment criteria, causing the main path to mistakenly turn into a side path. This method prioritizes calculating the degree of matching between each candidate direction and the displacement vector at the previous moment, and assigns the highest weight to the candidate direction that is most consistent with the historical direction. This allows the search operator to naturally traverse the intersection area along the original straight trajectory without being disturbed by vertical pipeline pixels. The candidate direction set explicitly excludes backtracking options that are completely opposite to the historical direction, further avoiding misjudgments of turning back at intersections. When there are valid pixels in the direction with the highest weight, the search immediately proceeds along that direction, and only when the straight direction is interrupted are other candidate directions evaluated in turn.

[0125] Specifically, the gap skipping mechanism enables the search process to perform cross-breakpoint logical reconnection of pixel discontinuities within a preset detection radius along the current motion vector direction. The specific steps are as follows:

[0126] When the topology search fails to detect any valid pipeline pixels within the preset search neighborhood, it is determined that there is a pixel discontinuity defect at the current search position, and gap skip detection is initiated, specifically as follows:

[0127] If there are no valid pipeline pixels in all candidate directions within the preset search neighborhood, it is determined to be a pixel discontinuity defect (such as dashed line intervals, scan breakpoints, etc.), the regular search is stopped, and gap jump detection is started. The jump direction is consistent with the pipeline extension direction at the previous moment to avoid deviating from the pipeline.

[0128] Pixel value detection is performed sequentially at each detection step position along the current motion vector direction within a preset detection radius, specifically as follows:

[0129] Based on the current search location Starting from the current motion vector direction, perform a unidirectional linear radar scan, with a detection step size of [missing information]. The value range is 2-30, and the core formula for pixel value detection at each detection step position is: ;

[0130] in, The direction component of the current motion vector. This is a function for reading pixel values ​​from a binary layer, used to read the corresponding coordinates. Pixel values;

[0131] Judgment after each test If the conditions are met, the pixel is considered a valid pipeline pixel, and the detection stops immediately. If the conditions are not met, the next detection step is performed until the number of steps reaches 30 or a valid pixel is detected.

[0132] If a non-zero valid pipeline pixel is detected within the preset detection radius, logical interpolation is used to complete the path from the current search position to the location of the valid pipeline pixel into a continuous path. Specifically:

[0133] If a valid pipeline pixel is detected at a certain step, a linear interpolation algorithm is used to complete the path between the current position and the pixel position, and the missing pixel is set as the valid pipeline pixel value to achieve reconnection across breakpoints;

[0134] The search position is then moved to the location where a valid pipeline pixel is detected, and the topology search continues, specifically as follows:

[0135] After path completion, the search position is switched to the effective pipeline pixel position, the normal topology search is resumed, and motion vector inertial constraints are enabled;

[0136] This process can be repeated to achieve continuous reconnection of dashed lines at multiple intervals.

[0137] The specific steps for performing pixel value detection sequentially at each detection step position within a preset detection radius along the current motion vector direction are as follows:

[0138] Starting from the current search position, the system moves pixel by pixel along the current motion vector direction, with the number of steps not exceeding the number of pixels corresponding to the preset detection radius. Specifically:

[0139] Starting from the current search position, the system moves pixel by pixel along the direction of motion, with the number of steps not exceeding the preset detection radius. If the detection position exceeds the range of the binarized layer, the system stops moving and is determined to have not detected any valid pipeline pixels.

[0140] After each step, obtain the pixel value at the current position and determine whether the pixel value is a non-zero valid pipeline pixel. Specifically:

[0141] After each step, the pixel value of that position in the binarized layer is obtained. Since the binarized layer has slight noise, as long as the pixel value is not 0, it is determined to be a valid pipeline pixel to avoid missing weak pipeline pixels.

[0142] If a pixel is determined to be a valid pipeline pixel, the stepping stops and the current position coordinates are recorded, specifically as follows:

[0143] If a valid pipeline pixel is detected at a certain step, stop stepping immediately, record the coordinates of that position and the number of steps, and mark that position as the "jump endpoint";

[0144] If no valid pipeline pixels are detected after the number of steps reaches the preset detection radius corresponding to the number of pixels, the current path is marked as a segment to be backtracked and the candidate direction is switched to continue detection. Specifically:

[0145] When the number of steps Reaching the preset detection radius (Right now If no valid pipeline pixels are found at any step position, it is determined that there is no continuous pipeline in the current direction, and the current path is marked as a segment to be backtracked. The marking format is as follows: ;

[0146] in, The coordinates of the starting point of the search for the current path. The coordinates of the current search position are recorded, along with the direction of movement of the segment to be backtracked. ;

[0147] Backtrack the search position to the previous valid search position. Switch candidate directions (excluding invalid directions), and detect from high to low weight until a valid pipeline pixel is found or all directions are invalid;

[0148] If all directions are ineffective, continue backtracking and repeating the operation to avoid getting stuck in an invalid path.

[0149] In this implementation scheme, the gap skipping mechanism gives the topology search the ability to actively detect areas where pixels are missing. When the conventional neighborhood search is difficult to continue due to the encounter with dotted line gaps or scanning breakpoints, this mechanism will not terminate the tracking process. Instead, it will perform a unidirectional radar scan within a preset distance along the original extension direction of the pipeline, actively searching for pipeline pixels on the opposite bank of the break. Once a valid pixel is recaptured within the scanning range, the blank area in the middle is filled into a continuous path through logical interpolation, and the search position is directly jumped to the opposite bank to continue tracking. This process can be triggered repeatedly, thereby completely connecting the dotted line signal loops that were originally physically broken, avoiding the dotted lines being incorrectly interpreted as unrelated discrete line segments. If no subsequent pixels are found within the scanning range, the current path is marked as a segment to be traced back and the candidate direction is switched to prevent excessive consumption of computing resources in invalid directions.

[0150] Specifically, the steps for generating the topological connection relationship between two elements based on the starting element and the target element encountered during the topology search are as follows:

[0151] When the probe mask corresponding to the search starting point touches the probe mask of another element during the topology search process, the unique identifiers of the starting element and the target element are obtained through the mapping relationship between pixel coordinates and element identifiers. Specifically:

[0152] During the topology search, the positional relationship between the search position and the detection mask of each element is monitored in real time. When the search position falls within the detection mask of two elements at the same time, it is determined that the two masks are touching.

[0153] Preset pixel coordinates and component identifier mapping table Each pixel within the element mask is mapped to a unique identifier for that element. ;

[0154] The two identifiers corresponding to the touch position are retrieved from the mapping table. The identifier that matches the element to which the search starting point belongs is the starting element identifier. The other is the target component identifier. ;

[0155] Component identifiers use the format "category prefix + serial number";

[0156] For example, resistor ,capacitance The serial numbers increase sequentially according to the detection order;

[0157] The unique identifiers of the starting element and the target element are recorded as a set of topology connection pairs, specifically:

[0158] In the order of "start-target", the identifiers of the two elements are recorded as topological connection pairs and associated with the corresponding search identifiers to distinguish the connection relationships of different paths;

[0159] To avoid duplicate records, if the same set of connection pairs already exists, only the path length is updated (the shortest path is taken), and key-value pairs are used for storage.

[0160] Calculate the search path length from the search start point to the touch position, and store the search path length as the path length attribute of the topology connection pair. Specifically:

[0161] The path length is calculated using the Euclidean distance accumulation method, and the formula is: ;

[0162] in, As the starting coordinates of the search, The coordinates of the touch location;

[0163] Since the search step size is 1 pixel, the distance between adjacent steps is 1 pixel during horizontal and vertical detection, and 1 pixel during diagonal detection. This can be simplified to counting the number of steps (when there are 4 neighborhoods, the number of steps is the path length).

[0164] After the calculation is complete, the path length will be... Retain two decimal places as the attribute of the join pair.

[0165] In this implementation scheme, when the search path arrives at the detection mask of another element from the starting element, the system quickly locks the unique identifiers of the two end elements through a preset pixel coordinate and element identifier mapping table, and records them as topological connection relationships in the form of ordered pairs of start and target. This process relies on the independent identity code assigned to each element to ensure that regardless of whether the element has a text tag, it can have a stable traceable identifier in the topological network, thereby solving the problem that unnamed elements are easily missed in logical association. For possible duplicate connection paths, this method automatically removes duplicates by comparing existing records and retaining the shortest path length, ensuring the simplicity and accuracy of the topological relationship table. The path length is calculated by accumulating Euclidean distance and stored as a connection attribute, providing a quantitative reference for pipeline length statistics, material estimation, or path impedance analysis.

[0166] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0167] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for digital parsing of industrial drawings based on multimodal visual fusion and topology search, characterized in that, Includes the following steps: Acquire existing industrial drawing images, perform target detection on the images, and output candidate bounding boxes containing component spatial location information and preliminary categories; Based on candidate bounding boxes, scale normalization and geometric transformation enhancement are performed on the local image regions corresponding to the candidate bounding boxes, and the normalized local image regions are output. The normalized local region of the image is input into the trained deep residual network, which extracts deep semantic features and performs fine-grained semantic re-arbitration, outputting the final category of the element. Based on the candidate bounding boxes, the pixels inside the element indicated by the candidate bounding boxes are zeroed on the binarized layer of the image, and a probe mask is built at the edge of the element to determine the search starting point. Starting from the search origin, a topology search with motion vector inertia constraint and gap jump mechanism is performed within the preset search neighborhood. The motion vector inertia constraint makes the search process prioritize maintaining the same motion direction as the displacement vector at the previous moment to avoid erroneous turning at pipeline intersections. The gap jump mechanism makes the search process perform cross-breakpoint logical reconnection of pixel discontinuous defects along the current motion vector direction within the preset detection radius until it touches the detection mask of another element. Based on the starting element and the target element touched during the topology search, a topological connection relationship between the two elements is generated.

2. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps for target detection in an image are as follows: The existing industrial drawings are resampled at high resolution and mapped to a preset detection space. A neural network based on an anchorless architecture is used to extract multi-level semantic features; Coordinate regression and preliminary class probability tests are performed on industrial components using a decoupled head structure; Output candidate bounding boxes and preliminary categories containing component spatial location information.

3. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps for scaling and geometric transformation enhancement of the local image region corresponding to the candidate bounding box are as follows: Map the local image region corresponding to the candidate bounding box to a tensor space with a preset normalized pixel size; Perform a fixed-angle rotation transformation on a local region of the scaled image; Perform a fixed-position alignment transformation on local regions of the scaled image; Complete the scale normalization and geometric transformation enhancement processing of local image regions.

4. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps for performing fine-grained semantic re-arbitration using the trained deep residual network are as follows: The local regions of the image, after scale normalization and geometric transformation enhancement, are input into the trained deep residual network; Deep semantic features of local regions in an image are extracted using a deep residual network; Calculate the probability distribution of component categories based on deep semantic features; The component category with the highest confidence level in the probability distribution is output as the final category of the component.

5. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps for constructing a probe mask at the edge of the component to determine the search starting point are as follows: After setting the pixels inside the element to zero based on the coordinate range of the candidate bounding box, the element edges are processed using geometric morphology operators. Construct a probe mask with a preset pixel width at the edge of the component; Perform a bitwise AND operation on the pipeline pixels in the probe mask and binarized layer to lock the precise intersection point between the pipeline and the component edge; The exact intersection point is determined as the starting point for the topology search.

6. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps of using motion vector inertial constraints to prioritize maintaining the same direction of motion as the displacement vector from the previous moment during the search process to avoid erroneous turns at pipeline intersections are as follows: The topology search is limited to an iterative exploration within a preset search neighborhood. In the initial search phase, the displacement vector is taken as the normal direction of the element edge, and the displacement vector at the previous moment is obtained by the subsequent detection step size. Calculate the degree of same-direction matching between each candidate detection direction and the displacement vector at the previous moment, and assign the candidate detection direction with the highest degree of same-direction matching the highest search weight; When there are valid pipeline pixels in the candidate probe direction corresponding to the highest search weight, perform the next long probe along that direction.

7. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 6, characterized in that, The degree of matching in the same direction is obtained by calculating the dot product of the candidate detection direction vector and the displacement vector at the previous moment. The vector dot product with the largest value corresponds to the highest degree of matching in the same direction. The specific steps are as follows: Obtain the displacement vector from the previous moment and represent it as a two-dimensional direction vector; Obtain the candidate detection direction vectors corresponding to each candidate direction within the preset search neighborhood of the current search position, and represent each candidate detection direction vector as a two-dimensional direction vector; Calculate the dot product of each candidate detection direction vector and the displacement vector at the previous moment; The candidate probe direction vector corresponding to the vector with the largest dot product is determined as the candidate probe direction vector with the highest degree of matching in the same direction.

8. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The gap skipping mechanism enables the search process to perform cross-breakpoint logical reconnection of pixel discontinuities within a preset detection radius along the current motion vector direction. The specific steps are as follows: When the topology search fails to detect any valid pipeline pixels within the preset search neighborhood, it is determined that there is a pixel discontinuity defect at the current search position, and gap skip detection is initiated. Pixel value detection is performed sequentially at each detection step position within a preset detection radius along the current motion vector direction; If a non-zero valid pipeline pixel is detected within the preset detection radius, the path between the current search position and the location of the valid pipeline pixel is completed into a continuous path through logical interpolation. Jump the search position to the location where a valid pipeline pixel is detected, and continue the topology search.

9. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 8, characterized in that, The specific steps for performing pixel value detection sequentially at each detection step position within a preset detection radius along the current motion vector direction are as follows: Starting from the current search position, move pixel by pixel along the current motion vector direction, with the number of steps not exceeding the number of pixels corresponding to the preset detection radius; After each step, obtain the pixel value at the current position and determine whether the pixel value is a non-zero valid pipeline pixel; If a pixel is determined to be a valid pipeline pixel, stop stepping and record the current position coordinates; If no valid pipeline pixels are detected after the number of steps reaches the preset detection radius corresponding to the number of pixels, the current path is marked as a segment to be backtracked and the candidate direction is switched to continue detection.

10. The method for digital parsing of industrial drawings based on multimodal visual fusion and topology search according to claim 1, characterized in that, The specific steps for generating the topological connection relationship between two elements based on the starting element and the target element encountered during the topology search are as follows: When the probe mask corresponding to the search starting point touches the probe mask of another element during the topology search process, the unique identifier of the starting element and the unique identifier of the target element are obtained through the mapping relationship between pixel coordinates and element identifiers. Record the unique identifier of the starting element and the unique identifier of the target element as a set of topology connection pairs; Calculate the search path length from the search start point to the touch position, and store the search path length as the path length attribute of the topology connection pair.

Images