A method and device for monitoring bridge construction progress based on image registration

Abstract

This invention provides a method and apparatus for monitoring bridge construction progress based on image registration. The method includes: extracting content features from a real-world image and style features from a BIM image, constructing a joint objective function, inputting the real-world image into a style transfer adaptor to obtain a BIM-stylized real-world image; inputting the real-world image into a dense matching model and parsing the output confidence matrix; introducing a segmentation mask for the target region and calculating the average matching confidence of the real-world image within the target region based on the segmentation mask; calculating the spatial gradient by performing tensor misalignment difference operations on the deformation field and constructing a deformation penalty term by combining a mask and truncation control mechanism; cross-weighting and fusing the positive and negative confidence excitations and deformation penalties to output the total score for each construction stage and determine the actual construction progress. This invention effectively solves the problems of interference from complex construction site backgrounds and forced mismatches, significantly improving the robustness of the measurement and the efficiency of automated supervision.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and engineering information technology, and more specifically, to a method and apparatus for monitoring bridge construction progress based on image registration. Background Technology

[0002] In the construction progress monitoring of large infrastructure such as bridges, the accuracy of image registration directly affects the reliability and automation level of progress assessment. Since adjacent construction stages are often highly similar in visual features (e.g., bridge decks or components with only a very small added structural proportion), traditional image matching methods based on global similarity or local feature descriptors (such as feature point alignment extracted by the SIFT algorithm) heavily rely on the distribution patterns of image textures, making it difficult to accurately capture subtle structural progress evolutions. While existing deep learning-based image feature matching networks (such as Transformer-based dense matching feature extraction models) have largely solved the robustness issues of large viewpoints and cross-modal applications, their matching index construction is mostly geared towards maximizing similarity. When processing the matching mapping between BIM model renderings and real-world images, they often ignore the physical spatial anomalies inherent in incorrect matching mapping relationships, thus failing to effectively differentiate the confidence scores between correct and incorrect matching stages, making it difficult to accurately determine construction progress.

[0003] Existing technologies still have the following shortcomings: 1) Global similarity scoring suffers from background interference. Image matching scores typically use the feature consistency of the entire image as a metric. In open scenes such as bridges, background elements like the sky and water occupy most of the pixels, making the area of ​​the target structure to be detected too small. This makes the global similarity score easily diluted by the background, resulting in difficulty in differentiating scores at different stages.

[0004] 2) Lack of penalty for deformation anomalies in mismatches. Most existing feature matching networks are geared towards maximizing feature similarity and overlap. In order to forcibly align non-existent building structures, the model causes severe local folding or stretching in the deformation field, making it easy for mismatches to obtain artificially high similarity scores.

[0005] 3) Significant differences exist between cross-domain images. There are huge texture differences between real construction scene photos and BIM images, and direct matching is prone to feature degradation and mismatch. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing a method and device for monitoring bridge construction progress based on image registration. By combining image segmentation, style transfer, and optimization of scoring algorithms, it solves the technical problems of interference from complex construction site backgrounds and forced mismatches.

[0007] According to a first aspect of the present invention, a method for monitoring bridge construction progress based on image registration is provided, comprising the following steps: Acquire real-world images of bridges, extract content features from real-world images and style features from BIM images using deep learning networks, construct a joint objective function, input the real-world images into a preset style transfer adapter, and obtain BIM-stylized real-world images. BIM-stylized realistic images are input into a dense matching model, and the output confidence matrix is ​​parsed. A segmentation mask for the target region is introduced, and the average matching confidence of image pairs within the target region is calculated based on the segmentation mask. Based on the dense deformation field output by the dense matching model, the local gradient of the deformation field is obtained through spatial difference operation, and the deformation index characterizing the local distortion amplitude is obtained to obtain the overall deformation degree. A scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. Extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, and cross-weight and fuse the average matching confidence of the positive and negative sides with the deformation penalty term to output the total score of each construction stage and determine the actual construction progress.

[0008] Based on the above technical solution, the present invention can also be improved as follows.

[0009] Optionally, the step of using a deep learning network to extract content features from real-world images and style features from BIM images, and constructing a joint objective function, includes: Using a pre-trained deep learning network with frozen weights, content feature maps of real-world images are extracted through deep convolutional layers, and style feature maps of BIM images are extracted through multi-scale convolutional layers. Calculate the mean squared error between the content feature maps of the generated image and the real image, and construct the content loss; Calculate the Gram matrix between the style feature maps of the generated image and the BIM image and obtain the mean squared error to construct the style loss; The content loss and the style loss are combined into a joint objective function according to preset weights.

[0010] Optionally, the construction of the joint objective function, which involves inputting the real-world image into a preset style transfer adaptor to obtain a BIM-stylized real-world image, includes: A joint objective function is constructed to iteratively differentiate and optimize the pixel matrix of the generated image at the pixel level, thereby achieving cross-modal feature fusion. Using a mask extracted from real images, an adaptive masking operation is performed on the invalid regions of the generated image after pixel-level optimization, constraining the style transfer range and maintaining the original geometric boundaries, resulting in a BIM-stylized real-world image that retains the spatial geometric boundaries.

[0011] Optionally, calculating the average matching confidence of the image pair within the target region based on the segmentation mask includes: Obtain the dense matching map and pixel-by-pixel matching confidence matrix output by the image matching model; Extract the set of valid matching points that legally fall into the target region using a segmentation mask; The confidence matrix is ​​masked by segmentation masking to remove interference from non-target background; The confidence scores of the effective matching pixels are accumulated and normalized to obtain the average matching confidence score of the target region.

[0012] Optionally, the step of accumulating and normalizing the confidence scores of valid matching pixels to obtain the average matching confidence score of the target region further includes: Perform tensor dimension adaptive alignment on the segmentation mask and count the total number of valid pixels within the mask; When the total number of valid pixels is greater than zero, the overlap matrix is ​​filtered by element-wise multiplication using the segmentation mask to force the confidence of non-target regions to zero. The confidence scores in the filtered overlap matrix are summed globally and divided by the total number of valid pixels to obtain the local average matching confidence. When the total number of valid foreground pixels is not greater than zero, an abnormal degradation mechanism is triggered, and the global mean of the original overlap matrix is ​​taken as the confidence score.

[0013] Optionally, the scaling factor is introduced to adjust the deformation index to ensure that the deformation index and the matching confidence level are consistent in numerical scale, resulting in a deformation penalty term including: In the horizontal and vertical directions, the spatial gradients in the horizontal and vertical directions are obtained by subtracting the misalignment difference between adjacent pixels and filling the constant boundary of the deformation field tensor. The L2 norms of the spatial gradients in the horizontal and vertical directions are calculated and summed to construct a global distortion map that quantifies the relative rate of change of spatial displacement. The average deformation value of the target region is obtained by filtering effective pixels and summing and averaging the global distortion map using a segmentation mask. The average deformation distortion value is multiplied by a preset scaling factor, and a nonlinear truncation function is applied to limit it to a preset upper threshold, outputting the final deformation penalty term.

[0014] Optionally, the cross-weighted fusion of the average matching confidence in both positive and negative directions with the deformation penalty term includes: The positive local average matching confidence is used as the incentive term, the positive deformation penalty term is used as the negative constraint, and the positive evaluation score is calculated by combining the set weight coefficients. The reverse local average matching confidence is used as the incentive term, the reverse deformation penalty term is used as the negative constraint, and the reverse evaluation score is calculated by combining the set weight coefficients. Based on the set symmetric balance weights, the positive evaluation scores and the negative evaluation scores are linearly weighted and fused to construct a bidirectional joint total score based on cycle consistency.

[0015] Optionally, the process for determining the actual construction progress includes: Traverse the BIM image sequence containing all construction stages and iteratively generate a two-way joint total score set of the real image relative to the BIM image of each stage; The total score set is sorted in descending order, and the BIM construction stage corresponding to the candidate with the highest total score is extracted as the judgment result. Extract the total score of the first-ranked team and the total score of the second-ranked team to calculate the confidence gap. If the confidence gap is less than a preset threshold, output a stage confusion risk warning.

[0016] According to a second aspect of the present invention, a bridge construction progress monitoring device based on image registration is provided, comprising: The cross-modal style transfer module is used to acquire real-world images of bridges, extract content features from real-world images and style features from BIM images using a deep learning network, construct a joint objective function, and input the real-world images into a preset style transfer adapter to obtain BIM-stylized real-world images. The local confidence aggregation module is used to input BIM-stylized realistic images into the dense matching model, parse the output confidence matrix, introduce the segmentation mask of the target region, and calculate the average matching confidence of image pairs in the target region based on the segmentation mask; The deformation penalty construction module is used to obtain the local gradient of the deformation field based on the dense deformation field output by the dense matching model through spatial difference operation, obtain the deformation index characterizing the local distortion amplitude, and obtain the overall deformation degree; a scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. The bidirectional fusion and progress determination module is used to extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, and perform cross-weighted fusion of the average matching confidence of the positive and negative sides and the deformation penalty term to output the total score of each construction stage and determine the actual construction progress.

[0017] According to a second aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the construction progress monitoring method based on image matching and bidirectional scoring as described in the first aspect.

[0018] The technical effects and advantages of this invention are as follows: This invention discloses a method and device for monitoring bridge construction progress based on image registration, and proposes an automated progress assessment strategy that integrates cross-modal style transfer and physical deformation constraints. First, in the image matching stage, this invention constructs a cascaded architecture consisting of a deep learning style transfer adaptor and a dense matching network: on the one hand, the style transfer network unifies the real image to the BIM style, effectively eliminating the modal barrier between the real construction lighting and materials and the BIM rendering, significantly reducing the difficulty of feature extraction; on the other hand, by introducing a target region segmentation mask, end-to-end automation of the feature matching point search and filtering process is achieved. This automated design allows end users to accurately eliminate environmental background interference and strictly focus the calculations on the core construction area without any tedious manual point selection operations in practical applications. Secondly, this invention introduces a deformation penalty mechanism based on deformation fields. By quantifying and analyzing local spatial gradients, it accurately captures the feature folding and stretching phenomena caused by forced matching when the matching model faces structural inconsistencies between two images, such as redundant components due to construction progress being ahead of or behind schedule. This transforms macroscopic structural differences into microscopic numerical penalty scores, solving the problem of forced mismatches that easily occur in traditional algorithms. Finally, this invention employs a bidirectional scoring weighted fusion mechanism, effectively filtering out occlusion interference and overfitting risks from a single perspective through cross-validation, thus improving the robustness of the scoring. This invention is applicable to scenarios such as smart construction sites, building lifecycle management, and automated engineering supervision. Through a fully automated matching and calculation process, it not only significantly reduces the usage threshold for end users, but its accuracy, anti-interference ability, and automation level in construction progress determination are also significantly superior to traditional manual visual verification and existing single image matching technologies. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1A flowchart illustrating an image registration-based bridge construction progress monitoring method provided in this embodiment of the invention. Figure 2 A flowchart illustrating the image registration-based bridge construction progress monitoring method provided in this embodiment of the invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To address the technical challenges in automated construction progress monitoring, such as significant modal differences between real-world images and BIM models, severe background interference, and the susceptibility to forced mismatches in complex structures, this invention proposes a construction progress monitoring strategy based on cross-modal image matching and bidirectional scoring. In model construction, this invention employs a cascaded architecture combining a deep learning style transfer adaptor and a dense matching network: on one hand, the style transfer network extracts style features from the BIM image and transfers them to the content features of the real-world image, generating a BIM-stylized real-world image that eliminates modal barriers; on the other hand, the processed image pairs are input into an image matching model to obtain dense matching features and deformation fields. In the scoring mechanism, a segmentation mask for the target region is introduced to shield against environmental background interference, automatically selecting valid matching points and calculating the average matching confidence within the target region; simultaneously, local spatial gradient analysis is performed on the deformation field to accurately identify feature folding and stretching phenomena caused by mismatches in construction components (such as ahead-of-schedule or behind-of-schedule progress), quantifying these as deformation penalty scores. The evaluation process employs a bidirectional matching mechanism (i.e., from real-world images to BIM images, and from BIM images to real-world images). The bidirectional matching confidence level is used as a positive incentive, and the bidirectional deformation penalty score is used as a negative penalty, resulting in a weighted fusion and a total score. Finally, by traversing the BIM images of each construction stage and selecting the one with the highest total score, the actual construction progress is automatically determined.

[0023] This invention is applicable to scenarios such as smart construction site construction, building lifecycle management (BIM / PLM) and automated engineering supervision. Through cross-modal style unification, automatic masking to resist interference, and a two-way scoring mechanism based on physical deformation constraints, it realizes end-to-end fully automatic evaluation of construction progress. Its monitoring accuracy, anti-occlusion robustness and implementation efficiency are significantly better than traditional manual visual verification and existing one-way image matching methods.

[0024] Understandably, given the deficiencies in the background technology, this invention proposes a method for monitoring bridge construction progress based on image registration, specifically as follows: Figure 1 As shown, it includes the following steps: S1: Obtain a real-world image of the bridge, use a deep learning network to extract the content features of the real-world image and the style features of the BIM image, construct a joint objective function, input the real-world image into a preset style transfer adapter, and obtain a BIM-stylized real-world image. In this embodiment, the real-world image is a photograph taken during bridge construction. The content feature is the spatial structure of the bridge. The style feature is the bridge's patterns and textures.

[0025] This invention is based on cross-modal feature extraction, pixel-level optimization and adaptive masking of deep learning networks. It inputs real images into a preset deep learning style transfer adaptor (such as VGG-19), extracts style features of the target BIM image and fuses them with the real image to generate a BIM stylized real image that eliminates modal differences.

[0026] The method of using deep learning networks to extract content features from real-world images and style features from BIM images to construct a joint objective function includes: Using a pre-trained deep learning network with frozen weights, content feature maps of real-world images are extracted through deep convolutional layers, and style feature maps of BIM images are extracted through multi-scale convolutional layers. Calculate the mean squared error between the content feature maps of the generated image and the real image, and construct the content loss; Calculate the Gram matrix between the style feature maps of the generated image and the BIM image and obtain the mean squared error to construct the style loss; The content loss and the style loss are combined into a joint objective function according to preset weights.

[0027] In one implementation, constructing a joint objective function and inputting the real-world image into a preset style transfer adaptor to obtain a BIM-stylized real-world image includes: A joint objective function is constructed to iteratively differentiate and optimize the pixel matrix of the generated image at the pixel level, thereby achieving cross-modal feature fusion. Using a mask extracted from the original real image, an adaptive masking operation is performed on the invalid region of the generated image after pixel-level optimization, and a BIM-styled real image with preserved spatial geometric boundaries is output.

[0028] Specifically, the processing steps of a deep learning style transfer adaptor include: Tensoring and scale-adaptive normalization are performed on real-world images and target BIM images; Using a deep learning network with frozen weights as a feature extractor, content features of real-world images are extracted through deep convolutional layers to preserve spatial structure, and style features of BIM images are extracted through multi-scale convolutional layers to capture texture and color. Calculate the mean squared error of the feature map to construct the content loss, and calculate the Gram matrix of the feature map to construct the style loss; The content loss and style loss are weighted and combined into a joint objective function, and the generated image is iteratively optimized at the pixel level to achieve the fusion of cross-modal features; By combining a two-dimensional Boolean mask extracted from the original real-world image, an adaptive masking operation is performed on the optimized image for invalid regions, resulting in a BIM-stylized real-world image that eliminates modal differences and strictly preserves spatial boundaries.

[0029] In the specific implementation process, the actual images taken at the construction site (defined as content images) and the corresponding BIM rendered images (defined as style images) are first acquired. Adaptive mapping and normalization are performed on the resolution of both, converting the RGB pixel matrix into a multidimensional tensor supported by the deep learning framework. Then, zero-mean normalization is performed channel-by-channel using preset statistical values ​​(such as ImageNet mean and standard deviation) to accelerate gradient convergence in subsequent high-dimensional spaces. Simultaneously, a generated image to be optimized is initialized.

[0030] In the feature extraction stage, a pre-constructed deep convolutional neural network (such as a truncated VGG-19 network) is used as the perceptual feature extractor, and all weight parameters of the network are frozen so that none of the weight parameters are updated during backpropagation. Real-world images, BIM images, and generated images are input into the network respectively. The system extracts feature maps from preset deep convolutional layers (such as conv_4) to represent spatial structure; and extracts feature maps from multiple preset shallow and mid-level convolutional layers (such as conv_1 to conv_5) to represent cross-scale texture and color distribution.

[0031] In constructing the joint objective function, to ensure that the generated image does not lose the physical geometry and topology of the real construction scene, the module calculates the generated image in the first... Layer feature representation Feature representation at the same layer as the original real-world image The mean squared error between them is defined as content loss. : (1) In the formula, This represents the calculated content loss value; This indicates a specified deep convolutional layer level in a pre-trained deep learning network; This indicates that the generated image to be optimized is at the 1st... In the layer feature map matrix, the first The first channel, the first Feature activation values ​​at spatial locations; The original reality image is represented in the first... In the layer feature map matrix, the first The first channel, the first Feature activation values ​​at spatial locations; where and These represent the channel dimension index and the spatial pixel dimension index of the feature map, respectively.

[0032] Simultaneously, to extract specific rendering styles from BIM images, the module introduces internal correlation calculations of feature maps. Within each style layer, the inner product between each channel of the feature map is calculated to generate a Gram matrix. If the... The feature map dimension of the layer is (Number of channels) multiplied by (Height-width product), then the Gram matrix The calculation is as follows: (2) In the formula, Indicates the first In the Gram matrix corresponding to the layer feature map, the first... Line 1 The element values ​​of the column are used to quantize the first... The first feature channel and the first Spatial correlation between feature channels; This represents the selected style extraction convolutional layer level in the pre-trained deep learning network; Indicates the first The total number of spatial pixels in the layer feature map, which is the product of the feature map height and width; This represents the pixel position index of the feature map in the unfolded spatial dimension; Indicates the image at the 1st In the layer feature map, the first The first channel in the Feature activation values ​​at spatial locations; Indicates the image at the 1st In the layer feature map, the first The first channel in the Feature activation values ​​at spatial locations; where and Indicates the indexes of different channels in the feature map.

[0033] The total style loss is constructed by calculating the mean square error between the Gram matrix of the generated image and the Gram matrix of the target BIM image, and by weighting and summing the errors across multiple layers. .

[0034] Subsequently, the content loss and style loss were weighted according to a preset weighting coefficient ( and By performing linear combinations, the final joint objective function is formed: (3) In the formula, This represents the total loss value of the joint objective function; This represents the preset content loss weighting coefficient; Indicates the content loss value; This represents the preset style loss weighting coefficient; This represents the style loss value.

[0035] At this stage, the system uses quasi-Newton methods (such as the L-BFGS algorithm) or momentum gradient descent to directly calculate the gradient of the joint objective function with respect to the generated image pixel matrix. After multiple pixel-level iterations of optimization, the total loss function converges, completing the cross-fusion of the feature space.

[0036] Finally, spatial consistency correction based on geometric masks is performed. Regions with extremely low or invalid pixel intensity in the original real-world image are extracted using a preset threshold, and a two-dimensional Boolean mask is constructed. After denormalizing the optimized generated image, this Boolean mask is used to perform an element-by-element masking operation, forcibly resetting pixels at invalid coordinate positions marked by the mask to zero or the original background color.

[0037] By employing a pre-defined style transfer adaptor, the cross-modal image matching problem (real photos and BIM renderings) is proactively reduced in dimensionality and transformed into a same-modal matching problem before being input into the matching model. This design effectively eliminates the interference of complex lighting, material reflections, and color deviations in the real construction environment on subsequent feature extraction, allowing the matching model to focus more on the topological structure and positional relationships of building components, significantly improving the accuracy and robustness of dense matching feature finding.

[0038] S2: Input the BIM-stylized realistic image into the dense matching model, parse the output confidence matrix; introduce the segmentation mask of the target region, and calculate the average matching confidence of the image pair in the target region based on the segmentation mask; In this embodiment, the processed real-world image and the corresponding BIM image are input into a dense matching model (such as RoMa v2), and the matching overlap matrix output by the model is parsed. Then, a segmentation mask is introduced, and after tensor dimension alignment and effective pixel statistics, the local average matching confidence that is strictly focused on the target building component is calculated through mask filtering.

[0039] The step of calculating the average matching confidence of the image pair within the target region based on the segmentation mask includes: Obtain the dense matching map and pixel-by-pixel matching confidence matrix output by the image matching model; Extract the set of valid matching points that legally fall within the target area using a mask; The confidence matrix is ​​masked using a mask to remove interference from the non-target background. The confidence scores of the effective matching pixels are accumulated and normalized to obtain the average matching confidence score of the target region.

[0040] In practice, dense matching models such as RoMa are used to find pixel-level correspondences in image alignment. However, actual construction site photos contain a large amount of dynamic background unrelated to the construction progress (such as the sky, clouds, moving tower cranes, temporary scaffolding, etc.). If the matching confidence of the entire image is directly calculated, the random matching scores of these backgrounds will severely dilute the matching weights of the core building components. Therefore, this embodiment introduces segmentation masks for accurate evaluation of local areas.

[0041] The step of accumulating and normalizing the confidence scores of valid matching pixels to obtain the average matching confidence score of the target region also includes: Perform tensor dimension adaptive alignment on the segmentation mask and count the total number of valid pixels within the mask; When the total number of valid pixels is greater than zero, the overlap matrix is ​​filtered by element-wise multiplication using the segmentation mask to force the confidence of non-target regions to zero. The confidence scores in the filtered overlap matrix are summed globally and divided by the total number of valid pixels to obtain the local average matching confidence. When the total number of valid foreground pixels is not greater than zero, an abnormal degradation mechanism is triggered, and the global mean of the original overlap matrix is ​​taken as the confidence score.

[0042] In the specific implementation process, firstly, preprocessed real-world construction images (as source images) and BIM rendered images of specific construction stages (as target images) are input into a pre-trained dense matching network for forward propagation calculation. The model outputs a dictionary of prediction results (preds) containing multi-dimensional features.

[0043] The system extracts an overlap matrix representing the pixel-level matching probability between two images, denoted as... Each element in this matrix The range is usually normalized to 1. Between these values, the magnitude directly reflects the coordinates in the source image. The confidence level of a pixel in the target image is that it has found a reliable source match. The closer the value is to 1, the more reliable the match; the closer the value is to 0, the more likely the pixel belongs to an occlusion or background with no corresponding relationship.

[0044] In actual construction sites, real-world images often contain a large amount of background interference that is not related to the main building. To accurately assess construction progress, this invention introduces a segmentation mask corresponding to the source image, denoted as . The mask is usually a binary matrix. Areas with a value of 1 in the mask represent the core building components that need to be monitored (such as bridge towers and bridge decks), while areas with a value of 0 represent irrelevant background.

[0045] Before performing matrix operations, device-level tensor synchronization is first performed to split the mask. Migration to overlap matrix On the same computing device, to ensure the efficiency of underlying computation. Subsequently, the system performs dynamic detection and adaptive dimensionality expansion operations on the tensor dimension of the mask: if the input mask is detected to be a two-dimensional tensor (i.e., shape...), then... The system adds a channel dimension at the end using the unsqueeze operation, forcing it to align into a three-dimensional tensor (shape: This ensures that the subsequent element-wise Hadamard product can be legally executed in the underlying computation graph through broadcasting.

[0046] After completing dimension alignment, the system performs a global summation operation on the segmentation mask to count the total number of valid foreground pixels in the current mask, denoted as the valid pixel scalar. The calculation formula is as follows: (4) In the formula, This scalar represents the total number of valid foreground pixels obtained from the statistics. This represents the total pixel width of the mask in the horizontal direction; This represents the total pixel height of the mask in the vertical direction; and These represent the horizontal and vertical coordinate indices of the mask in two-dimensional space, respectively. A mask representing a real image in spatial coordinates The pixel value at the given coordinate is 1 when the coordinate belongs to the valid target area and 0 when it belongs to the non-target background area.

[0047] By judgment The mask validity is checked using the numerical value: when When the current mask is deemed valid and a valid foreground target exists, the process enters the mask-guided local aggregation calculation branch; if... If the mask fails to load, it is determined that the mask is invalid, triggering the exception degradation handling mechanism.

[0048] When the mask verification is valid ( When using a segmentation mask For the overlap matrix A mask filtering operation is performed. Specifically, the two are multiplied element-wise, forcing the confidence score of non-target regions (mask value of 0) to zero, retaining only the matching confidence score of the core building component regions. Then, the filtered confidence matrix is ​​globally summed and divided by the total number of valid pixels calculated using the previously listed formula. This allows for the calculation of the local average matching confidence score that is strictly focused on the foreground target. The mathematical expression for this process is: (5) In the formula, This represents the local average matching confidence score focused on the target region, calculated after masking filtering. and These represent the pixel dimensions in the horizontal and vertical directions, respectively. and These represent the horizontal and vertical coordinate indices of the matrix in two-dimensional space, respectively. This indicates that in the overlap matrix output by the dense matching model, the values ​​located at... Match confidence at the location; Represents the real-world image mask in coordinates Binary pixel distribution at the location; This represents the total number of valid pixels within the mask area, calculated in advance.

[0049] This implementation process highly condenses the millions of pixel-level matching probabilities scattered across the entire image into a macroscopic scalar score that characterizes the degree of matching between the current real-world image and the BIM model at that stage. Because the dilution effect of irrelevant backgrounds is completely eliminated, this score can extremely accurately reflect the alignment of the main construction progress.

[0050] As a guarantee of system robustness, when a mask failure is detected (i.e., no mask is provided or...), In order to ensure uninterrupted monitoring, the system will automatically degrade to a global calculation mode. At this time, the system directly processes the original overlap matrix. The average of all elements is used as the global mean as the matching confidence score for the image pair. (6) In the formula, This represents the global average matching confidence score after the abnormal degradation mechanism is triggered; and These represent the total pixel dimensions of the original overlap matrix in the horizontal and vertical directions, respectively. and These represent the horizontal and vertical coordinate indices in two-dimensional space, respectively. This represents the coordinates in the original overlap matrix output by the dense matching model. The confidence level of the match at the given location.

[0051] The aforementioned degradation mechanism ensures system availability under conditions of mask generation failure or extreme input, thereby improving the stability of the entire progress monitoring device's engineering implementation.

[0052] S3: Based on the dense deformation field output by the dense matching model, the local gradient of the deformation field is obtained through spatial difference operation, and the deformation index characterizing the local distortion amplitude is obtained to obtain the overall deformation degree; a scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. In this embodiment, the deformation field tensor output by the matching model is extracted, and the spatial gradient features of the deformation field are calculated by performing misalignment difference operations on adjacent pixels in the horizontal and vertical directions. Then, the degree of local abnormal deformation in the effective building component area is quantified by combining the segmentation mask, and a scaling factor and upper limit threshold are introduced for truncation. Finally, the deformation penalty term is constructed and output.

[0053] The introduction of a scaling factor to adjust the deformation index ensures that the deformation index and the matching confidence level are consistent in numerical scale, resulting in a deformation penalty term including: In the horizontal and vertical directions, the spatial gradients in the horizontal and vertical directions are obtained by subtracting the misalignment difference between adjacent pixels and filling the constant boundary of the deformation field tensor. The L2 norms of the spatial gradients in the horizontal and vertical directions are calculated and summed to construct a global distortion map that quantifies the relative rate of change of spatial displacement. The average deformation value of the target region is obtained by filtering effective pixels and summing and averaging the global distortion map using a segmentation mask. The average deformation distortion value is multiplied by a preset scaling factor, and a nonlinear truncation function is applied to limit it to a preset upper threshold, outputting the final deformation penalty term.

[0054] The core of this implementation step lies in quantifying the forced deformation anomaly (Distortion / Tearing Anomaly) during the image matching process. In actual construction monitoring, real-world images may contain extra components built ahead of schedule or parts not yet built, while BIM renderings present an ideal state at a certain stage. To forcibly align these redundant or missing features, the matching network will generate severe stretching or folding in the deformation field of that area. Step S3 precisely captures these abnormal deformations by calculating the spatial partial derivative of the deformation field, thus providing a scientific basis for negative penalties in the overall score.

[0055] In a preferred embodiment, the dense deformation field is output based on the dense matching model, the spatial gradient is calculated by performing tensor misalignment difference operations on the dense deformation field, and a deformation penalty term is constructed by combining masking and truncation control mechanisms, including the following sub-steps: From the forward propagation prediction results of the dense matching model, the deformation field tensor of the forward matching (from the real image to the BIM image) is extracted and denoted as . .in, and Spatial resolution of the image, number of channels Typically, this is 2, representing the horizontal and vertical displacements of each pixel when mapped to the target image, respectively. This tensor macroscopically records the absolute spatial displacement experienced by each pixel in the source image in order to fit the target image.

[0056] To assess the severity of deformation, we need to analyze the relative displacement differences (i.e., spatial gradients) between adjacent pixels, rather than the absolute displacement of individual pixels. If adjacent pixels were originally close together but are now significantly separated after mapping, it indicates severe local structural tearing. The horizontal spatial gradient is calculated using efficient tensor misalignment slicing and subtraction operations. Spatial gradient in the vertical direction First, in the horizontal direction, the deformation field tensor is shifted one pixel to the left and subtracted from the original tensor. This is done to maintain strict consistency in the tensor's size across the spatial dimensions (i.e., to maintain...). (Regarding the shape), the system employs a constant padding mechanism, using the value 0 to fill in the missing right edge columns. The calculation formula is as follows: (7) Similarly, in the vertical direction, the deformation field tensor is shifted upwards by one pixel, and the difference is subtracted. Then, the missing rows at the bottom are padded with a constant value of 0 to obtain the vertical spatial gradient. (8) In the formula, In the deformation field tensor, the coordinates located in space are... The horizontal spatial gradient at that location In the deformation field tensor, the coordinates located in space are... The spatial gradient in the vertical direction at that location; This indicates that in the positive deformation field tensor output by the dense matching model, the coordinates are... The absolute spatial displacement eigenvector at the location; Through the tensor-level parallel computation described above, the system obtains the discrete first-order spatial partial derivatives of the deformable field with extremely high efficiency while avoiding time-consuming loop traversal.

[0057] In obtaining directional gradient and Next, the system needs to aggregate the differences in displacement vectors from multiple channels into a single scalar distortion level. For each pixel, the system calculates the L2 norm of its horizontal gradient vector and the L2 norm of its vertical gradient vector, and adds them together to generate a global distortion map that incorporates two-dimensional deformation features, denoted as . The mathematical expression for this process is: (9) In the formula, This indicates the generated global distortion map in spatial coordinates. The degree of distortion at the location; This indicates that the L2 norm (i.e., Euclidean norm) of a vector feature is calculated. In the deformation field tensor, the coordinates located in space are... The horizontal spatial gradient vector at that location; In the deformation field tensor, the coordinates located in space are... The spatial gradient vector in the vertical direction at that location; and These represent the horizontal and vertical coordinate indices of the global distortion map in two-dimensional space, respectively.

[0058] Distortion image The values ​​directly reflect the degree of structural tearing in that local area; the higher the value, the more abnormal the deformation.

[0059] Since the entire image may contain a large number of natural deformations unrelated to the background (such as erroneous optical flow caused by cloud drift), the system again introduces the real-image-specific segmentation mask from step S2. and total number of effective pixels The system utilizes a mask. Global distortion map Element-wise multiplication filtering is performed to mask deformation noise in areas outside the main building structure. Then, the filtered effective distortion values ​​are globally summed and divided by [the factor]. The average deformation distortion value was obtained by strictly focusing on the core construction target. : (10) In the formula, This represents the average deformation and distortion value of the target after masking. and These represent the pixel dimensions of the global distortion map and the mask in the horizontal and vertical directions, respectively. and These represent the horizontal and vertical coordinate indices of the matrix in two-dimensional space, respectively. Represents the global distortion map in spatial coordinates The degree of scalar distortion at the location; The mask representing the real image in coordinates Binarized pixel distribution at the location; This represents the total number of valid foreground pixels within the mask area, calculated from pre-statistics.

[0060] If the total number of valid pixels is detected If the mask fails, an abnormal degradation mechanism is triggered, and the unfiltered global distortion image is directly retrieved. The mean as .

[0061] Finally, the physical distortion values ​​need to be transformed into a reasonable penalty term that can be used for the final total score calculation. Directly subtracting the distortion mean might result in excessive penalties due to individual extreme noise points (such as abrupt matching changes caused by localized reflections), thereby completely destroying the stability of the matching score. Therefore, the system introduces a linear scaling factor. To amplify minute structural tearing features; simultaneously, a penalty upper limit threshold is introduced. A truncation control mechanism is constructed. The system calculates the scaled distortion value and applies a nonlinear truncation function (Min Function) to output the final deformation tearing penalty score. : (11) This truncation design provides the system with strong fault tolerance, ensuring that even under extreme conditions such as severe occlusion or abnormal optical flow estimation, a single deformation penalty will not be amplified indefinitely, thus preventing the scoring system from collapsing.

[0062] S4: Extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, cross-weight and fuse the average matching confidence of the positive and negative sides with the deformation penalty term, output the total score of each construction stage and determine the actual construction progress.

[0063] In this embodiment, the bidirectional overlap matrix and deformation field output by the matching model are obtained in parallel; the source image segmentation mask and BIM-specific mask are introduced independently, and the local matching confidence and deformation penalty in the forward and reverse directions are calculated in parallel; the bidirectional indicators are weighted and fused to obtain the total score of this stage, and the optimal construction progress stage corresponding to the real image is determined by a global traversal and descending sorting mechanism.

[0064] The method of cross-weighting and fusing the average matching confidence of both positive and negative sides with the deformation penalty term includes: The positive local average matching confidence is used as the incentive term, the positive deformation penalty term is used as the negative constraint, and the positive evaluation score is calculated by combining the set weight coefficients. The reverse local average matching confidence is used as the incentive term, the reverse deformation penalty term is used as the negative constraint, and the reverse evaluation score is calculated by combining the set weight coefficients. Based on the set symmetric balance weights, the positive evaluation scores and the negative evaluation scores are linearly weighted and fused to construct a bidirectional joint total score based on cycle consistency.

[0065] Specifically, the core of this step lies in leveraging the feature reuse characteristics of the deep matching network's underlying layer to achieve bidirectional cross-validation based on cycle consistency without increasing additional inference computing power, and outputting the final engineering progress conclusion.

[0066] The process for determining the actual construction progress includes: Traverse the BIM image sequence containing all construction stages and iteratively generate a two-way joint total score set of the real image relative to the BIM image of each stage; The total score set is sorted in descending order, and the BIM construction stage corresponding to the candidate with the highest total score is extracted as the judgment result. Extract the total score of the first-ranked team and the total score of the second-ranked team to calculate the confidence gap. If the confidence gap is less than a preset threshold, output a stage confusion risk warning.

[0067] In a preferred embodiment, the cross-weighted fusion of the positive and negative confidence incentives and deformation penalties includes the following sub-steps: The pre-processed actual construction drawing (denoted as...) ) and BIM renderings of specific construction phases (denoted as The input pairs are fed into a pre-trained dense matching network for a single forward inference. Thanks to the bidirectional attention or dual analytic mechanism within the matching network, the model can output four core tensors in parallel from its prediction dictionary (preds) after a single operation: the positive overlap matrix... positive deformation field and the reverse overlap matrix Reverse deformation field .

[0068] After obtaining the bidirectional feature tensor, the system loads two independent mask filtering paths and parallel index quantization: In the forward evaluation path, the system introduces a segmentation mask specific to real-world images. Reuse the logic in steps S2 and S3 to and By performing filtering and gradient differentiation, the positive local average matching confidence score is calculated. With positive deformation tearing penalty score .

[0069] In the reverse evaluation path, the system introduces a dedicated multi-class segmentation mask of the current BIM image in parallel. Using this BIM mask to... and Independent effective pixel filtering, local confidence aggregation, and misalignment differential distortion calculation are performed to derive the reverse local average matching confidence in parallel. With reverse deformation tearing penalty score .

[0070] To eliminate the one-way fitting bias caused by occlusion in the real scene, the system constructs a parameterized two-way fusion recipe. First, the one-way comprehensive score and the positive evaluation score are calculated separately. Reverse evaluation score Subsequently, according to the preset symmetrical balance weights (such as taking equal values), The positive and negative scores are linearly weighted and fused to calculate the bidirectional joint total score corresponding to the current BIM stage for the real-world image. : (12) Using real-world construction images as a benchmark, the system iterates through a pre-defined database containing BIM image sequences across all construction stages, repeatedly executing the matching and scoring process described above to generate a candidate score set for each real-world image relative to each BIM stage. Subsequently, the system uses a two-way joint total score... The candidate set is sorted in descending order. The candidate with the highest total score (i.e., ranked first) is extracted, and its corresponding BIM construction stage is output as the actual construction progress determination result of the current real-world image. Further, the system extracts the highest score (ranked first) and the second highest score (ranked second), and calculates the confidence gap between them (e.g., using the formula...). If the gap is greater than a preset threshold, the current progress matching result is considered reliable; if the gap is small, it indicates the risk of progress confusion between similar stages, thus providing a scientific decision-making reference for the automated supervision system.

[0071] Figure 2 This is a flowchart illustrating the construction progress monitoring method based on image matching and bidirectional scoring. The method receives external input images of the construction site and corresponding BIM model diagrams for each stage from different perspectives. A preprocessing module performs tensor quantization and dimensional alignment. Specifically, a pre-defined cross-modal style transfer module eliminates modal barriers between images. The processed image pairs are then input into a dense matching network for single forward inference, simultaneously extracting the bidirectional overlap matrix and deformation field. Subsequently, a dedicated segmentation mask is used, and a matrix mask filtering mechanism is employed to extract the confidence level of the target region. A deformation penalty term is constructed based on the spatial gradient of the deformation field. Finally, a bidirectional cross-fusion mechanism is used to output the total score for that stage.

[0072] In summary, the construction progress monitoring method based on image matching and bidirectional scoring proposed in this invention systematically introduces deformation field spatial gradient analysis (deformation penalty) and cyclic consistency (bidirectional symmetric evaluation) into cross-modal building image matching. Through tensor misalignment difference operations, mask-guided local confidence aggregation, and anomaly degradation control design, the model ensures extremely high stability and robustness in the face of complex and ever-changing dynamic construction site environments, effectively solving the problem of unidirectional forced mismatch caused by redundant or missing components. This model significantly improves the performance of automated calculation strategies and outperforms existing unidirectional visual verification methods in diverse engineering scenarios such as smart construction site supervision and BIM full lifecycle management, verifying its high industrial application value.

[0073] This invention integrates physical deformation constraints and bidirectional feature extraction, organically combining cross-modal alignment, mask filtering, single-pass forward inference, and a weighted penalty loss mechanism, demonstrating significant advantages in practical applications across multiple fields. The core penalty mechanism of this invention utilizes the microscopic displacement partial derivatives of the deformation field to accurately identify and quantify deformation tearing phenomena that violate physical structures. In practical applications, this model significantly improves the accuracy of progress determination under various occlusion conditions, greatly reducing the visual comparison costs and subjective errors of manual on-site supervision, and enhancing the response speed of digital project progress tracking.

[0074] As shown in the table below, the method of the present invention is significantly better than the existing comparative methods in terms of core indicators such as progress determination accuracy and score lead.

[0075]

[0076] This embodiment also discloses a construction progress monitoring device based on image matching and two-way scoring, including: The cross-modal style transfer module is used to acquire real-world images of bridges, extract content features from real-world images and style features from BIM images using a deep learning network, construct a joint objective function, and input the real-world images into a preset style transfer adapter to obtain BIM-stylized real-world images. The local confidence aggregation module is used to input BIM-stylized realistic images into the dense matching model, parse the output confidence matrix, introduce the segmentation mask of the target region, and calculate the average matching confidence of image pairs in the target region based on the segmentation mask; The deformation penalty construction module is used to obtain the local gradient of the deformation field based on the dense deformation field output by the dense matching model through spatial difference operation, obtain the deformation index characterizing the local distortion amplitude, and obtain the overall deformation degree; a scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. The bidirectional fusion and progress determination module is used to extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, and perform cross-weighted fusion of the average matching confidence of the positive and negative sides and the deformation penalty term to output the total score of each construction stage and determine the actual construction progress.

[0077] In addition, this invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described construction progress monitoring method based on image matching and bidirectional scoring.

[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0079] 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.

[0080] 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.

[0081] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring bridge construction progress based on image registration, characterized in that, include: Acquire real-world images of bridges, extract content features from real-world images and style features from BIM images using deep learning networks, construct a joint objective function, input the real-world images into a preset style transfer adapter, and obtain BIM-stylized real-world images. The BIM-stylized real-world image is input into the dense matching model, and the output confidence matrix is ​​parsed. A segmentation mask for the target region is introduced, and the average matching confidence of the real-world image within the target region is calculated based on the segmentation mask. Based on the dense deformation field output by the dense matching model, the local gradient of the deformation field is obtained through spatial difference operation, and the deformation index characterizing the local distortion amplitude is obtained to obtain the overall deformation degree. A scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. Extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, and cross-weight and fuse the average matching confidence of the positive and negative sides with the deformation penalty term to output the total score of each construction stage and determine the actual construction progress.

2. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The method of using deep learning networks to extract content features from real-world images and style features from BIM images, and constructing a joint objective function, includes: Using a pre-trained deep learning network with frozen weights, content feature maps of real-world images are extracted through deep convolutional layers, and style feature maps of BIM images are extracted through multi-scale convolutional layers. Calculate the mean squared error between the content feature maps of the generated image and the real image, and construct the content loss; Calculate the Gram matrix between the style feature maps of the generated image and the BIM image and obtain the mean squared error to construct the style loss; The content loss and the style loss are combined into a joint objective function according to preset weights.

3. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The construction of the joint objective function, which involves inputting the real-world image into a preset style transfer adaptor to obtain a BIM-stylized real-world image, includes: A joint objective function is constructed to iteratively differentiate and optimize the pixel matrix of the generated image at the pixel level, thereby achieving cross-modal feature fusion. Using a mask extracted from real images, an adaptive masking operation is performed on the invalid regions of the generated image after pixel-level optimization, constraining the style transfer range and maintaining the original geometric boundaries, resulting in a BIM-stylized real-world image that retains the spatial geometric boundaries.

4. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The step of calculating the average matching confidence of the image pair within the target region based on the segmentation mask includes: Obtain the dense matching map and pixel-by-pixel matching confidence matrix output by the image matching model; Extract the set of valid matching points that legally fall into the target region using a segmentation mask; The confidence matrix is ​​masked by segmentation masking to remove interference from non-target background; The confidence scores of the effective matching pixels are accumulated and normalized to obtain the average matching confidence score of the target region.

5. The method for monitoring bridge construction progress based on image registration according to claim 4, characterized in that, The step of accumulating and normalizing the confidence scores of valid matching pixels to obtain the average matching confidence score of the target region also includes: Perform tensor dimension adaptive alignment on the segmentation mask and count the total number of valid pixels within the mask; When the total number of valid pixels is greater than zero, the overlap matrix is ​​filtered by element-wise multiplication using the segmentation mask to force the confidence of non-target regions to zero. The confidence scores in the filtered overlap matrix are summed globally and divided by the total number of valid pixels to obtain the local average matching confidence. When the total number of valid foreground pixels is not greater than zero, an abnormal degradation mechanism is triggered, and the global mean of the original overlap matrix is ​​taken as the confidence score.

6. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The introduction of a scaling factor to adjust the deformation index ensures that the deformation index and the matching confidence level are consistent in numerical scale, resulting in a deformation penalty term including: In the horizontal and vertical directions, the spatial gradients in the horizontal and vertical directions are obtained by subtracting the misalignment difference between adjacent pixels and filling the constant boundary of the deformation field tensor. The L2 norms of the spatial gradients in the horizontal and vertical directions are calculated and summed to construct a global distortion map that quantifies the relative rate of change of spatial displacement. The average deformation value of the target region is obtained by filtering effective pixels and summing and averaging the global distortion map using a segmentation mask. The average deformation distortion value is multiplied by a preset scaling factor, and a nonlinear truncation function is applied to limit it to a preset upper threshold, outputting the final deformation penalty term.

7. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The method of cross-weighting and fusing the average matching confidence of both positive and negative sides with the deformation penalty term includes: The positive local average matching confidence is used as the incentive term, the positive deformation penalty term is used as the negative constraint, and the positive evaluation score is calculated by combining the set weight coefficients. The reverse local average matching confidence is used as the incentive term, the reverse deformation penalty term is used as the negative constraint, and the reverse evaluation score is calculated by combining the set weight coefficients. Based on the set symmetric balance weights, the positive evaluation scores and the negative evaluation scores are linearly weighted and fused to construct a bidirectional joint total score based on cycle consistency.

8. The method for monitoring bridge construction progress based on image registration according to claim 1, characterized in that, The process for determining the actual construction progress includes: Traverse the BIM image sequence containing all construction stages and iteratively generate a two-way joint total score set of the real image relative to the BIM image of each stage; The total score set is sorted in descending order, and the BIM construction stage corresponding to the candidate with the highest total score is extracted as the judgment result. Extract the total score of the first-ranked team and the total score of the second-ranked team to calculate the confidence gap. If the confidence gap is less than a preset threshold, output a stage confusion risk warning.

9. A bridge construction progress monitoring device based on image registration, characterized in that, include: The cross-modal style transfer module is used to acquire real-world images of bridges, extract content features from real-world images and style features from BIM images using a deep learning network, construct a joint objective function, and input the real-world images into a preset style transfer adapter to obtain BIM-stylized real-world images. The local confidence aggregation module is used to input BIM-stylized realistic images into the dense matching model, parse the output confidence matrix, introduce the segmentation mask of the target region, and calculate the average matching confidence of image pairs in the target region based on the segmentation mask; The deformation penalty construction module is used to obtain the local gradient of the deformation field based on the dense deformation field output by the dense matching model through spatial difference operation, obtain the deformation index characterizing the local distortion amplitude, and obtain the overall deformation degree; a scaling factor is introduced to adjust the scale of the deformation index so that the deformation index and the matching confidence are consistent in numerical scale, and the deformation penalty term is obtained. The bidirectional fusion and progress determination module is used to extract the inverse confidence matrix and inverse deformation field tensor output by the dense matching model, perform evaluation by combining the segmentation mask of the BIM image, and perform cross-weighted fusion of the average matching confidence of the positive and negative sides and the deformation penalty term to output the total score of each construction stage and determine the actual construction progress.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the construction progress monitoring method based on image matching and two-way scoring as described in any one of claims 1 to 8.

Images