Large cantilever bent cap pier construction supervision system in limited space

Abstract

The invention discloses a large cantilever bent cap pier construction supervision system in a limited space. The system comprises an image acquisition module, an image enhancement module, a construction supervision model establishment module and a large cantilever bent cap pier construction supervision module in the limited space. The invention belongs to the technical field of construction monitoring, and particularly relates to a large cantilever bent cap pier construction supervision system in a limited space, and the robustness of a model is enhanced by cascading a plurality of sub-networks; the triple attention unit extracts information of channels and spatial dimensions; absolute error loss and perception loss are combined; finally, the construction supervision accuracy is improved; calculating inter-class scores through the minimum distance; a query feature graph is reconstructed through ridge regression, and features of a query image and a support set image are matched more accurately; based on channel re-weighting and category allocation, the classification precision of the model is improved; the center loss is introduced to further optimize the representation of the features and improve the classification accuracy; and the efficient performance of the model in the construction supervision task is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of construction monitoring and processing, and in particular to a construction supervision system for a large cantilever cap beam bridge pier in a limited space. Background Art

[0002] The cantilever cap beam bridge pier construction supervision system is a comprehensive monitoring and management system used for real-time monitoring, quality control and progress management of all links in the construction process of large cantilever cap beam bridge piers. The system usually integrates a variety of technical means, such as image processing, sensor data acquisition, data analysis and artificial intelligence, with the aim of ensuring construction quality, improving construction efficiency and ensuring construction safety. However, the general construction supervision system has the problem of poor quality of original construction images, inability to accurately represent the details and structure of the image during optimization, poor ability to describe the details of the features, and thus low accuracy of the final construction monitoring; the general construction supervision system has the problem of being unable to capture the subtle features of each category, unable to effectively distinguish different categories, and unable to comprehensively measure the importance of each monitoring hazard level, which leads to poor monitoring results. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a large cantilever cap beam bridge pier construction supervision system in a limited space. The general construction supervision system has the problem that the original construction image quality is poor, the details and structure of the image cannot be accurately represented during optimization, and the details and description capabilities of the features are poor, which leads to low accuracy of the final construction monitoring. This scheme avoids the gradient vanishing problem by cascading multiple sub-networks, and at the same time enhances the robustness of the model and reduces the risk of overfitting by combining feature extraction at different levels; by constructing multi-level sub-networks and using the context modeling module, the global context of the image can be better understood; the information extraction module helps to extract detailed features of the image and enhance the feature expression capability; the triple attention unit effectively extracts information in the channel and spatial dimensions, thereby improving the details and description capabilities of the features; combined with the absolute error loss and the perceptual loss, it can optimize at both the pixel level and the feature level, thereby obtaining higher quality image enhancement. Effect; ultimately improve the accuracy of construction supervision and processing; in view of the fact that general construction supervision systems are unable to capture the subtle features of each category, cannot effectively distinguish different categories, and cannot comprehensively measure the importance of each monitored hazard level, which leads to poor monitoring results, this solution calculates the intra-class score of each channel to accurately represent the typical features of each category and reduce the impact of noise on the intra-class score; calculates the inter-class score through the minimum distance, which helps to enhance the sensitivity to inter-class differences; reconstructs the query feature map through ridge regression to more accurately match the features of the query image and the support set image, thereby improving the accuracy of recognition; based on channel reweighting and category assignment, more detailed classification results are obtained, and the classification accuracy of the model is increased; the introduction of center loss can further optimize the representation of features, making samples of the same category closer in the feature space and samples of different categories more separated, thereby improving the accuracy of classification; ensuring the efficient performance of the model in construction supervision tasks.

[0004] The technical solution adopted by the present invention is as follows: A construction supervision system for a large cantilever cap beam bridge pier in a limited space provided by the present invention comprises an image acquisition module, an image enhancement module, a construction supervision model establishment module and a construction supervision module for a large cantilever cap beam bridge pier in a limited space;

[0005] The image acquisition module acquires historical construction images;

[0006] The image enhancement module finally realizes the construction of the image enhancement module by designing a sub-network cascade mechanism, constructing a multi-level sub-network, establishing a triple attention unit and designing a loss function;

[0007] The construction supervision model building module calculates the intra-class score of each channel, calculates the inter-class score through the minimum distance, reconstructs the query feature map through ridge regression, obtains the classification result based on channel reweighting and category assignment, introduces the central loss to construct the loss function of the construction supervision model, and finally realizes the establishment of the construction supervision model;

[0008] The large cantilever cap beam bridge pier construction supervision module in the limited space realizes the recognition of the construction images collected in real time based on the established construction supervision model.

[0009] Furthermore, in the image acquisition module, the historical construction images are construction images of various links in the construction process of the historical large cantilever cap beam bridge piers; the historical construction images are annotated, and the annotation type is the hazard assessment level; and the hazard assessment level is used as an image label.

[0010] Furthermore, the image enhancement module includes a sub-network cascade mechanism design unit, a multi-level sub-network construction unit, a triple attention unit and a loss function design unit; specifically includes the following contents:

[0011] Sub-network cascade mechanism design unit; connect multiple sub-networks through the cascade mechanism to prevent gradient vanishing and overfitting, and improve the robustness and generalization ability of the network; the sub-network cascade mechanism includes a shallow feature extractor for extracting shallow features of the input image, i.e., the historical construction image; a deep feature extractor for extracting deep features after multi-level sub-network processing; an adaptive residual block, including multi-level sub-networks and triple attention units; the sub-network cascade mechanism process is expressed as follows:

[0012] ;

[0013] In the formula, is the output graph; It is the original historical construction image; is the weight matrix of the shallow feature extractor; It is the adaptive residual block of layers 1 to 3; It is the triple attention unit from layer 1 to layer 3; It is a deep feature extractor;

[0014] Multi-level sub-network construction unit; through the multi-level sub-network structure, rich low-resolution context features are obtained while retaining high-resolution features to provide accurate spatial representation; the multi-level sub-network includes residual attention blocks and selective kernel feature fusion; the residual attention block contains context modeling modules and information extraction modules, and the selective kernel feature fusion dynamically adjusts the receptive field for feature fusion; the process includes context modeling, information extraction and feature fusion; context modeling is expressed as: ;in, is the output feature after context modeling; is the input feature; is the weight matrix used to generate context representation; is the weight matrix used to calculate the attention weight; p is the number of features, j and m are the feature indices; and They are the jth feature and the mth feature respectively; information extraction is expressed as: ; Where R is the output feature after information extraction; F is the input feature map; It is a triple attention unit; It is context modeling; is the weight matrix of the convolutional layer; feature fusion is expressed as: ; Among them, U is the output feature after feature fusion; L1 and L2 are input feature maps of different scales; is global average pooling; is the fusion weight matrix; is the Sigmoid activation function; is the ReLU activation function;

[0015] Triple attention unit; by focusing on channel and spatial information to refine detailed features, the effect of feature extraction and enhancement is improved; the triple attention unit includes a pooling layer and three branches; the pooling layer performs maximum pooling and average pooling in the channel dimension, and the three branches establish interactions in different dimensions respectively; the process is expressed as: ;in, is the feature representation after pooling, and X is the input feature; is the maximum pooling; is average pooling; It is to concatenate the results of maximum pooling and average pooling;

[0016] Loss function design unit; the loss function of the image enhancement module includes absolute error loss and perceptual loss; the absolute error loss is used to calculate the pixel-level difference between the enhanced image and the real image; the perceptual loss is used to calculate the feature difference; the absolute error loss function is expressed as: ;in, is the absolute error loss function; is the enhanced image; is a real image; is a small constant; the perceptual loss function is expressed as: ;in, is the perceptual loss function; H and W are the height and width of the feature map respectively; is the value of the enhanced image feature map at position (i1, j1); is the value of the real image feature map at position (i1, j1); the loss function used by the final image enhancement module is expressed as: ; Where L is the final loss function used; and is the weight parameter; is the square of the L2 norm.

[0017] Furthermore, the construction supervision model building module includes a data set division unit, a task weight calculation unit, a feature map reconstruction unit, a channel reweighting unit, a category allocation unit and a loss function design unit; specifically includes the following contents:

[0018] Dataset division unit; randomly extract K samples from the original data set for each category to build a support set; for each selected category, randomly extract q samples from the remaining samples to build a query set;

[0019] Task weight calculation unit; specifically includes:

[0020] Calculate the intra-class score; the prototype P of each class n is the average of all instances in the support set in the nth class; the intra-class score vector Contains the intra-class scores of all channels; the intra-class score of the cth channel is expressed as:

[0021] ;

[0022] Where H and W are the height and width of the images in the support set and query set in the spatial dimension respectively; is the prototype vector of the cth channel in the nth class; is the channel average of the nth class; is the square of the L2 norm;

[0023] Calculate the channel average using the following formula:

[0024] ;

[0025] Where C is the total number of channels;

[0026] Calculate inter-class scores; introduce minimum distance, inter-class score vector Contains the inter-class scores of all channels; the inter-class score of the cth channel is expressed as:

[0027] ;

[0028] In the formula, N is the total number of categories; is the channel average of the lth class;

[0029] Calculate the task-related support weight; the formula used is as follows:

[0030] ;

[0031] In the formula, is the task-related support weight of the nth class; is the weight coefficient; is the score mapping between classes;

[0032] Calculate the intra-class score of the query image using the following formula:

[0033] ;

[0034] ;

[0035] In the formula, is the intra-class score of the query image of the cth channel; is the weight to measure the importance of the query image; is the score mapping within the query image class; is the intra-class score vector of the query image; M Q is the channel average of the query image;

[0036] Calculate the final task-related weights using the following formula:

[0037] ;

[0038] In the formula, is the final task-related weight; is the weight coefficient;

[0039] Feature map reconstruction unit: Reconstruct the query feature map through ridge regression, expressed as: ;in, is the weight matrix obtained by solving; Q is the feature map of the query image; is the regularization parameter in ridge regression; is the support set feature map of the nth class; W is the transformation matrix that maps the support set feature map to the query feature map; the query image reconstructed by the nth class is finally expressed as: ; is the reconstructed query image;

[0040] Channel reweighting unit; based on the task-related weight vector b n Reweight the channels, expressed as:

[0041] ;

[0042] ;

[0043] In the formula, It is the re-weighted feature map of the n-th query image; is the result after reweighting; is a matrix of all 1s; T is the transpose operation; It is an element-by-element multiplication, that is, weighting each channel of the feature map;

[0044] Category assignment unit: calculates the probability that the query image belongs to the nth category, expressed as: ; ;in, is the j2th query image x j2 The probability of belonging to the nth class; γ is a hyperparameter used to adjust the influence of distance; is a distance metric; and They are the feature maps of the j2-th query image and the reconstructed query image respectively; and The categories are A query image and a feature map of the reconstructed query image;

[0045] Loss function design unit; introducing the central loss, the loss function of the construction supervision model is expressed as: ; Where Lz is the loss function of the construction supervision model; is the query image Belong to The probability of the class; Nq is the total number of samples in the query set; and They are the nth and The feature matrix of the class; is the weight coefficient of the center loss; is the query image x j2 Features; is the query image x j2 The central feature of the category to which it belongs; the original data set is divided into a test set and a training set in advance and a threshold is set. When the model converges on the training set loss and the prediction accuracy of the test set is higher than the threshold, the construction supervision module for large cantilever cap beam piers in limited space is established.

[0046] Furthermore, the construction supervision module for large cantilever cap beam piers in limited space is based on an established construction supervision model; construction images are collected in real time and input into the construction supervision model, and a danger threshold is pre-set. When the danger assessment level output by the model is higher than the danger threshold, an early warning is given.

[0047] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0048] (1) In view of the problem that the original construction images of general construction supervision systems are of poor quality, the details and structures of the images cannot be accurately represented during optimization, and the details and description capabilities of the features are poor, which leads to low accuracy of the final construction monitoring. This scheme avoids the gradient vanishing problem by cascading multiple sub-networks, and enhances the robustness of the model by combining feature extraction at different levels to reduce the risk of overfitting; by constructing multi-level sub-networks and using the context modeling module, the global context of the image can be better understood; the information extraction module helps to extract detailed features of the image and enhance the feature expression capability; the triple attention unit effectively extracts information in the channel and spatial dimensions, thereby improving the details and description capabilities of the features; combined with the absolute error loss and perceptual loss, it can optimize at both the pixel level and the feature level, thereby obtaining a higher quality image enhancement effect; and ultimately improve the accuracy of construction supervision and processing.

[0049] (2) In view of the fact that general construction supervision systems are unable to capture the subtle features of each category, cannot effectively distinguish different categories, and cannot comprehensively measure the importance of each monitored hazard level, which leads to poor monitoring results, this scheme calculates the intra-class score of each channel to accurately represent the typical features of each category and reduce the impact of noise on the intra-class score; calculates the inter-class score by the minimum distance, which helps to enhance the sensitivity to the difference between categories; reconstructs the query feature map through ridge regression to more accurately match the features of the query image and the support set image, thereby improving the recognition accuracy; obtains more detailed classification results based on channel reweighting and category assignment, and increases the classification accuracy of the model; introduces the center loss to further optimize the representation of features, so that samples of the same category are closer in the feature space and samples of different categories are more separated, thereby improving the classification accuracy; and ensures the efficient performance of the model in construction supervision tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a flow chart of a large cantilever cap beam bridge pier construction supervision system in a limited space provided by the present invention;

[0051] Figure 2 It is a flowchart of the image enhancement module;

[0052] Figure 3 Schematic diagram of the process of building modules for the construction supervision model.

[0053] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0055] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0056] Example 1, see Figure 1 The present invention provides a large cantilever cap beam bridge pier construction supervision system in a limited space, including an image acquisition module, an image enhancement module, a construction supervision model establishment module and a large cantilever cap beam bridge pier construction supervision module in a limited space;

[0057] The image acquisition module acquires historical construction images; and sends the data to the image enhancement module;

[0058] The image enhancement module receives the data sent by the image acquisition module; finally realizes the construction of the image enhancement module by designing a sub-network cascade mechanism, constructing a multi-level sub-network, establishing a triple attention unit and designing a loss function; and sends the data to the construction supervision model establishment module;

[0059] The construction supervision model establishment module receives the data sent by the image enhancement module; calculates the intra-class score of each channel, calculates the inter-class score by the minimum distance, reconstructs the query feature map by ridge regression, obtains the classification result based on the channel reweighting and category assignment, introduces the center loss to construct the loss function of the construction supervision model, and finally realizes the establishment of the construction supervision model; and sends the data to the large cantilever cap beam bridge pier construction supervision module in a limited space;

[0060] The large cantilever cap beam bridge pier construction supervision module in the limited space receives the data sent by the construction supervision model establishment module; and realizes the recognition of the construction images collected in real time based on the established construction supervision model.

[0061] Example 2, see Figure 1This embodiment is based on the above embodiment. In the image acquisition module, the historical construction images are construction images of various links in the construction process of the historical large cantilever cap beam bridge pier; the historical construction images are annotated, and the annotation type is the hazard assessment level; the hazard assessment level is used as the image label.

[0062] Example 3, see Figure 1 and Figure 2 , this embodiment is based on the above embodiment, the image enhancement module includes a sub-network cascade mechanism design unit, a multi-level sub-network construction unit, a triple attention unit and a loss function design unit; specifically includes the following contents:

[0063] Sub-network cascade mechanism design unit; connect multiple sub-networks through the cascade mechanism to prevent gradient vanishing and overfitting, and improve the robustness and generalization ability of the network; the sub-network cascade mechanism includes a shallow feature extractor for extracting shallow features of the input image, i.e., the historical construction image; a deep feature extractor for extracting deep features after multi-level sub-network processing; an adaptive residual block, including multi-level sub-networks and triple attention units; the sub-network cascade mechanism process is expressed as follows:

[0064] ;

[0065] In the formula, is the output graph; It is the original historical construction image; is the weight matrix of the shallow feature extractor; It is the adaptive residual block of layers 1 to 3; It is the triple attention unit from layer 1 to layer 3; It is a deep feature extractor;

[0066] Multi-level sub-network construction unit; through the multi-level sub-network structure, rich low-resolution context features are obtained while retaining high-resolution features to provide accurate spatial representation; the multi-level sub-network includes residual attention blocks and selective kernel feature fusion; the residual attention block contains context modeling modules and information extraction modules, and the selective kernel feature fusion dynamically adjusts the receptive field for feature fusion; the process includes context modeling, information extraction and feature fusion; context modeling is expressed as: ;in, is the output feature after context modeling; is the input feature; is the weight matrix used to generate context representation; is the weight matrix used to calculate the attention weight; p is the number of features, j and m are the feature indices; and They are the jth feature and the mth feature respectively; information extraction is expressed as: ; Where R is the output feature after information extraction; F is the input feature map; It is a triple attention unit; It is context modeling; is the weight matrix of the convolutional layer; feature fusion is expressed as: ; Among them, U is the output feature after feature fusion; L1 and L2 are input feature maps of different scales; is global average pooling; is the fusion weight matrix; is the Sigmoid activation function; is the ReLU activation function;

[0067] Triple attention unit; by focusing on channel and spatial information to refine detailed features, the effect of feature extraction and enhancement is improved; the triple attention unit includes a pooling layer and three branches; the pooling layer performs maximum pooling and average pooling in the channel dimension, and the three branches establish interactions in different dimensions respectively; the process is expressed as: ;in, is the feature representation after pooling, and X is the input feature; is the maximum pooling; is average pooling; It is to concatenate the results of maximum pooling and average pooling;

[0068] Loss function design unit; the loss function of the image enhancement module includes absolute error loss and perceptual loss; the absolute error loss is used to calculate the pixel-level difference between the enhanced image and the real image; the perceptual loss is used to calculate the feature difference; the absolute error loss function is expressed as: ;in, is the absolute error loss function; is the enhanced image; is a real image; is a small constant; the perceptual loss function is expressed as: ;in, is the perceptual loss function; H and W are the height and width of the feature map respectively; is the value of the enhanced image feature map at position (i1, j1); is the value of the real image feature map at position (i1, j1); the loss function used by the final image enhancement module is expressed as: ; Where L is the final loss function used; and is the weight parameter; is the square of the L2 norm.

[0069] By performing the above operations, the general construction supervision system has the problem of poor quality of original construction images, inability to accurately represent the details and structure of the image during optimization, poor ability to describe the details and features, and thus low accuracy of the final construction monitoring. This solution avoids the gradient vanishing problem by cascading multiple sub-networks, and enhances the robustness of the model by combining feature extraction at different levels to reduce the risk of overfitting; by constructing multi-level sub-networks and using the context modeling module, the global context of the image can be better understood; the information extraction module helps to extract detailed features of the image and enhance the feature expression capability; the triple attention unit effectively extracts information in the channel and spatial dimensions, thereby improving the details and description capabilities of the features; combined with the absolute error loss and the perceptual loss, it can optimize at both the pixel level and the feature level to obtain higher quality image enhancement effects; ultimately improving the accuracy of construction supervision and processing.

[0070] Example 4, see Figure 1 and Figure 3 , this embodiment is based on the above embodiment, and the construction supervision model establishment module includes a data set division unit, a task weight calculation unit, a feature map reconstruction unit, a channel reweighting unit, a category allocation unit and a loss function design unit; specifically includes the following contents:

[0071] Dataset division unit; randomly extract K samples from the original data set for each category to build a support set; for each selected category, randomly extract q samples from the remaining samples to build a query set;

[0072] Task weight calculation unit; specifically includes:

[0073] Calculate the intra-class score; the prototype P of each class n is the average of all instances in the support set in the nth class; the intra-class score vector Contains the intra-class scores of all channels; the intra-class score of the cth channel is expressed as:

[0074] ;

[0075] Where H and W are the height and width of the images in the support set and query set in the spatial dimension respectively; is the prototype vector of the cth channel in the nth class; is the channel average of the nth class; is the square of the L2 norm;

[0076] Calculate the channel average using the following formula:

[0077] ;

[0078] Where C is the total number of channels;

[0079] Calculate inter-class scores; introduce minimum distance, inter-class score vector Contains the inter-class scores of all channels; the inter-class score of the cth channel is expressed as:

[0080] ;

[0081] In the formula, N is the total number of categories; is the channel average of the lth class;

[0082] Calculate the task-related support weight; the formula used is as follows:

[0083] ;

[0084] In the formula, is the task-related support weight of the nth class; is the weight coefficient; is the score mapping between classes;

[0085] Calculate the intra-class score of the query image using the following formula:

[0086] ;

[0087] ;

[0088] In the formula, is the intra-class score of the query image of the cth channel; is the weight to measure the importance of the query image; is the score mapping within the query image class; is the intra-class score vector of the query image; M Q is the channel average of the query image;

[0089] Calculate the final task-related weights using the following formula:

[0090] ;

[0091] In the formula, is the final task-related weight; is the weight coefficient;

[0092] Feature map reconstruction unit: Reconstruct the query feature map through ridge regression, expressed as: ;in, is the weight matrix obtained by solving; Q is the feature map of the query image; is the regularization parameter in ridge regression; is the support set feature map of the nth class; W is the transformation matrix that maps the support set feature map to the query feature map; the query image reconstructed by the nth class is finally expressed as: ; is the reconstructed query image;

[0093] Channel reweighting unit; based on the task-related weight vector b n Reweight the channels, expressed as:

[0094] ;

[0095] ;

[0096] In the formula, It is the re-weighted feature map of the n-th query image; is the result after reweighting; is a matrix of all 1s; T is the transpose operation; It is an element-by-element multiplication, that is, weighting each channel of the feature map;

[0097] Category assignment unit: calculates the probability that the query image belongs to the nth category, expressed as: ; ;in, is the j2th query image x j2 The probability of belonging to the nth class; γ is a hyperparameter used to adjust the influence of distance; is a distance metric; and They are the feature maps of the j2-th query image and the reconstructed query image respectively; and The categories are A query image and a feature map of the reconstructed query image;

[0098] Loss function design unit; introducing the central loss, the loss function of the construction supervision model is expressed as: ; Where Lz is the loss function of the construction supervision model; is the query image Belong to The probability of the class; Nq is the total number of samples in the query set; and They are the nth and The feature matrix of the class; is the weight coefficient of the center loss; is the query image x j2 Features; is the query image x j2 The central feature of the category to which it belongs; the original data set is divided into a test set and a training set in advance and a threshold is set. When the model converges on the training set loss and the prediction accuracy of the test set is higher than the threshold, the construction supervision module for large cantilever cap beam piers in limited space is established.

[0099] By performing the above operations, the general construction supervision system cannot capture the subtle features of each category, cannot effectively distinguish different categories, and cannot comprehensively measure the importance of each monitored hazard level, which leads to poor monitoring results. This solution calculates the intra-class score of each channel to accurately represent the typical features of each category and reduce the impact of noise on the intra-class score; calculates the inter-class score by the minimum distance, which helps to enhance the sensitivity to the difference between categories; reconstructs the query feature map through ridge regression to more accurately match the features of the query image and the support set image, thereby improving the recognition accuracy; obtains more detailed classification results based on channel reweighting and category assignment, and increases the classification accuracy of the model; introduces the center loss to further optimize the representation of features, so that samples of the same category are closer in the feature space, and samples of different categories are more separated, thereby improving the classification accuracy; and ensures the efficient performance of the model in construction supervision tasks.

[0100] Example 5, see Figure 1 This embodiment is based on the above embodiment. The construction supervision module for large cantilever cap beam piers in limited space is based on an established construction supervision model. Construction images are collected in real time and input into the construction supervision model, and the monitoring hazard level based on the model output is used as the identification result.

[0101] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0102] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0103] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A construction supervision system for large cantilever cap beam bridge piers in limited space, characterized by: The system includes an image acquisition module, an image enhancement module, a construction supervision model building module, and a large cantilever cap beam bridge pier construction supervision module in a limited space; The image acquisition module acquires historical construction images; The image enhancement module finally realizes the construction of the image enhancement module by designing a sub-network cascade mechanism, constructing a multi-level sub-network, establishing a triple attention unit and designing a loss function; The construction supervision model building module calculates the intra-class score of each channel, calculates the inter-class score through the minimum distance, reconstructs the query feature map through ridge regression, obtains the classification result based on channel reweighting and category assignment, introduces the central loss to construct the loss function of the construction supervision model, and finally realizes the establishment of the construction supervision model; The large cantilever cap beam bridge pier construction supervision module in the limited space realizes the recognition of the real-time collected construction images based on the established construction supervision model; The image enhancement module contains a sub-network cascade mechanism design unit; the sub-network cascade mechanism includes a shallow feature extractor for extracting shallow features of an input image, i.e., a historical construction image; a deep feature extractor for extracting deep features after multi-level sub-network processing; an adaptive residual block, including a multi-level sub-network and a triple attention unit; the sub-network cascade mechanism process is represented as follows: ； In the formula, is the output graph; It is the original historical construction image; is the weight matrix of the shallow feature extractor; It is the adaptive residual block of layers 1 to 3; It is the triple attention unit from layer 1 to layer 3; It is a deep feature extractor.

2. According to the construction supervision system of a large cantilever cap beam bridge pier in a limited space in claim 1, it is characterized by: The image enhancement module includes a sub-network cascade mechanism design unit, a multi-level sub-network construction unit, a triple attention unit and a loss function design unit; specifically includes the following contents: Sub-network cascading mechanism design unit; Multi-level sub-network building unit; Multi-level sub-network includes residual attention block and selective kernel feature fusion; Residual attention block includes context modeling module and information extraction module, and selective kernel feature fusion dynamically adjusts the receptive field for feature fusion; The process includes context modeling, information extraction and feature fusion; Context modeling is expressed as: ;in, is the output feature after context modeling; is the input feature; is the weight matrix used to generate context representation; is the weight matrix used to calculate the attention weight; p is the number of features, j and m are the feature indices; and They are the jth feature and the mth feature respectively; information extraction is expressed as: ; Where R is the output feature after information extraction; F is the input feature map; It is a triple attention unit; It is context modeling; is the weight matrix of the convolutional layer; feature fusion is expressed as: ; Among them, U is the output feature after feature fusion; L1 and L2 are input feature maps of different scales; is global average pooling; is the fusion weight matrix; is the Sigmoid activation function; is the ReLU activation function; Triple attention unit; The triple attention unit includes a pooling layer and three branches; The pooling layer performs maximum pooling and average pooling in the channel dimension, and the three branches establish interactions in different dimensions respectively; The process is expressed as: ;in, is the feature representation after pooling, and X is the input feature; is the maximum pooling; is average pooling; It is to concatenate the results of maximum pooling and average pooling; Loss function design unit; the loss function of the image enhancement module includes absolute error loss and perceptual loss; the absolute error loss is used to calculate the pixel-level difference between the enhanced image and the real image; the perceptual loss is used to calculate the feature difference; the absolute error loss function is expressed as: ;in, is the absolute error loss function; is the enhanced image; is a real image; is a small constant; the perceptual loss function is expressed as: ;in, is the perceptual loss function; H and W are the height and width of the feature map respectively; is the value of the enhanced image feature map at position (i1, j1); is the value of the real image feature map at position (i1, j1); the loss function used by the final image enhancement module is expressed as: ; Where L is the final loss function used; and is the weight parameter; is the square of the L2 norm.

3. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 2 is characterized by: The construction supervision model establishment module includes a data set division unit, a task weight calculation unit, a feature map reconstruction unit, a channel reweighting unit, a category allocation unit and a loss function design unit; Includes the following: Dataset division unit; randomly extract K samples from the original data set for each category to build a support set; for each selected category, randomly extract q samples from the remaining samples to build a query set; Task weight calculation unit; Feature map reconstruction unit: Reconstruct the query feature map through ridge regression, expressed as: ;in, is the weight matrix obtained by solving; Q is the feature map of the query image; is the regularization parameter in ridge regression; is the support set feature map of the nth class; W is the transformation matrix that maps the support set feature map to the query feature map; the query image reconstructed by the nth class is finally expressed as: ; is the reconstructed query image; Channel reweighting unit; based on the task-related weight vector b n Reweight the channels, expressed as: ； ； In the formula, It is the re-weighted feature map of the n-th query image; is the result after reweighting; is a matrix of all 1s; T is the transpose operation; It is an element-by-element multiplication, that is, weighting each channel of the feature map; Category allocation unit; Loss function design unit.

4. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 3 is characterized by: The task weight calculation unit specifically includes: Calculate the intra-class score; the prototype P of each class n is the average of all instances in the support set in the nth class; the intra-class score vector Contains the intra-class scores of all channels; the intra-class score of the cth channel is expressed as: ； Where H and W are the height and width of the images in the support set and query set in the spatial dimension respectively; is the prototype vector of the cth channel in the nth class; is the channel average of the nth class; is the square of the L2 norm; Calculate the channel average using the following formula: ； Where C is the total number of channels; Calculate inter-class scores; introduce minimum distance, inter-class score vector Contains the inter-class scores of all channels; the inter-class score of the cth channel is expressed as: ； In the formula, N is the total number of categories; is the channel average of the lth class; Calculate the task-related support weight; the formula used is as follows: ； In the formula, is the task-related support weight of the nth class; is the weight coefficient; is the score mapping between classes; Calculate the intra-class score of the query image using the following formula: ； ； In the formula, is the intra-class score of the query image of the cth channel; is the weight to measure the importance of the query image; is the score mapping within the query image class; is the intra-class score vector of the query image; M Q is the channel average of the query image; Calculate the final task-related weights using the following formula: ； In the formula, is the final task-related weight; is the weight coefficient.

5. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 4 is characterized by: The category assignment unit specifically calculates the probability that the query image belongs to the nth category, which is expressed as: ; ;in, is the j2th query image x j2 The probability of belonging to the nth class; γ is a hyperparameter used to adjust the influence of distance; is a distance metric; and They are the feature maps of the j2-th query image and the reconstructed query image respectively; and The categories are The query image and the feature map of the reconstructed query image.

6. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 5 is characterized by: The loss function design unit specifically introduces the central loss, and the loss function of the construction supervision model is expressed as: ; Where Lz is the loss function of the construction supervision model; is the query image Belong to The probability of the class; Nq is the total number of samples in the query set; and They are the nth and The feature matrix of the class; is the weight coefficient of the center loss; is the query image x j2 Features; is the query image x j2 The central feature of the category to which it belongs; the original data set is divided into a test set and a training set in advance and a threshold is set. When the model converges on the training set loss and the prediction accuracy of the test set is higher than the threshold, the construction supervision module for large cantilever cap beam piers in limited space is established.

7. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 6 is characterized by: In the image acquisition module, the historical construction images are construction images of various links in the construction process of the historical large cantilever cap beam bridge pier; the historical construction images are annotated, and the annotation type is the hazard assessment level; the hazard assessment level is used as an image label.

8. The construction supervision system for large cantilever cap beam bridge piers in limited space according to claim 7 is characterized by: The construction supervision module for large cantilever cap beam bridge piers in limited space is based on an established construction supervision model; construction images are collected in real time and input into the construction supervision model, and a danger threshold is pre-set. When the danger assessment level output by the model is higher than the danger threshold, an early warning is given.

Images