Design drawing auditing method and device and storage medium

Abstract

The invention provides a design drawing auditing method and device and a storage medium. The method comprises the following steps: training an encoder adapted to a plurality of subsections and a plurality of deep detection networks adapted to a single subsection; the method comprises the following steps: receiving a three-dimensional design drawing constructed for a building when digital delivery is carried out on the constructional engineering; in the design drawing, target image data is intercepted from object models of various engineering objects under the multiple subsections; inputting the target image data into an encoder to extract common image features used for dividing a plurality of segments; for the same branch, inputting the public image features into a depth detection network to identify a target depth value of the object model; grading the depth of the object model according to the target depth value; generating a quality value about depth for the design drawing according to the grade; and generating auditing information about digital delivery for the design drawing according to the quality value. According to the embodiment, the efficiency of checking the depth of the object model in the design drawing is effectively improved.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical field of computer vision, and in particular, relate to a design drawing review method, device, and storage medium. Background Art

[0002] During the design phase of a construction project, designers construct various design drawings of the building in the Building Information Modeling (BIM). After completion, they export them into GDB format and upload them to the drawing review platform for review by profession to complete the BIM drawing review and delivery.

[0003] The review officer assigns tasks to appropriate reviewers and verifiers based on the review project information. The reviewer conducts a preliminary review of the professional drawings. When the preliminary review is completed, the drawings are handed over to the verifier for further review. When the review is completed, the review officer will be informed of the opinions, and the review officer will reply to the designer.

[0004] During the self-examination, preliminary examination and review process, it is necessary to review whether the depth of the design drawings meets the requirements in accordance with the design specifications.

[0005] In this process, the depth of design drawings is mainly reviewed manually, which has high labor costs and fuzzy boundaries of depth. It is easy to make mistakes due to subjectivity, resulting in low review efficiency. Summary of the invention

[0006] In view of this, an embodiment of the present application provides a design drawing review method, device and storage medium to improve the efficiency of the depth review of design drawings of construction projects.

[0007] A first aspect of an embodiment of the present application provides a design drawing review method, which is applied to a building information modeling platform, and the method includes:

[0008] Training encoders adapted to multiple parts and multiple deep detection networks adapted to a single part;

[0009] Receive constructed, three-dimensional design drawings of the building when digitally delivering the construction project;

[0010] In the design drawing, target image data is captured for object models of various engineering objects under a plurality of divisions;

[0011] Inputting the target image data into the encoder to extract common image features for dividing a plurality of sub-parts;

[0012] For the same sub-portion, inputting the common image features into the depth detection network to identify a target depth value of the object model;

[0013] classifying the object model in depth according to the target depth value;

[0014] generating a quality value of depth for the design drawing according to the grade;

[0015] Audit information on digital delivery is generated for the design drawings according to the quality value.

[0016] A second aspect of an embodiment of the present application provides a design drawing review device, which is applied to a building information modeling platform, and the device includes:

[0017] A model training module, used to train encoders adapted to multiple divisions and multiple deep detection networks adapted to a single division;

[0018] A design drawing receiving module, used for receiving three-dimensional design drawings of buildings when digitally delivering construction projects;

[0019] A target image data interception module is used to intercept target image data for object models of various engineering objects under multiple divisions in the design drawings;

[0020] A common image feature extraction module, used for inputting the target image data into the encoder to extract common image features for dividing the multiple sub-parts;

[0021] A target depth value recognition module, for inputting the common image features into the depth detection network to recognize a target depth value of the object model for the same sub-portion;

[0022] A depth level division module, used for dividing the object model into levels in depth according to the target depth value;

[0023] A quality value generating module, used for generating a quality value about depth for the design drawing according to the grade;

[0024] An audit information generation module is used to generate audit information about digital delivery for the design drawings based on the quality value.

[0025] A third aspect of an embodiment of the present application provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for reviewing design drawings as described in the first aspect above is implemented.

[0026] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for reviewing design drawings as described in the first aspect above is implemented.

[0027] A fifth aspect of an embodiment of the present application provides a computer program product. When the computer program product is run on a computer, the computer executes the design drawing review method described in the first aspect.

[0028] In this embodiment, encoders adapted to multiple divisions and multiple depth detection networks adapted to a single division are trained; when digital delivery of a construction project is performed, a three-dimensional design drawing of a building is received; in the design drawing, target image data is intercepted for the object model of various engineering objects under multiple divisions; the target image data is input into the encoder to extract the common image features used to divide the multiple divisions; for the same division, the common image features are input into the depth detection network to identify the target depth value of the object model; the object model is graded in depth according to the target depth value; a quality value about the depth is generated for the design drawing according to the grade; and audit information about digital delivery is generated for the design drawing according to the quality value. This embodiment provides automated depth auditing in the building information model platform, provides reference for self-examination, preliminary examination, re-examination and other auditing processes, reduces the reliance on manual auditing of design drawings, greatly reduces labor costs, and, based on deep learning to determine the level of depth, it can reduce the influence of subjectivity of manual auditing, provide a standardized scale, and improve the accuracy of the level of depth, thereby effectively improving the efficiency of auditing the depth of object models in design drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 It is a schematic diagram of a design drawing review method provided in an embodiment of the present application;

[0031] Figure 2 is an example diagram of depth provided by an embodiment of the present application;

[0032] Figure 3 It is a schematic diagram of the structure of an encoder and a depth detection network provided in an embodiment of the present application;

[0033] Figure 4 It is a schematic diagram of a design drawing review device provided in an embodiment of the present application;

[0034] Figure 5 It is a schematic diagram of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0035] In the following description, specific details such as specific system structures, technologies, etc. are proposed for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from hindering the description of the present application.

[0036] The technical solution of the present application is described below through specific embodiments.

[0037] Reference Figure 1 , shows a schematic diagram of a design drawing review method provided by an embodiment of the present application. This embodiment can be applied to a building information modeling platform. An auxiliary plug-in is installed on the client of the BIM platform, and the plug-in is used to complete the review of the design drawings. Specifically, the following steps may be included:

[0038] Step 101: train encoders adapted to multiple divisions and multiple deep detection networks adapted to single divisions.

[0039] Generally speaking, the entire construction project will be divided into multiple sections, each section will be divided into multiple sub-sections, and each sub-section will be divided into multiple engineering objects, so as to construct design drawings for the engineering objects.

[0040] For example, in a certain specification, construction projects are divided into foundation and base, main structure, building decoration and decoration, building roof, building water supply, drainage and heating, ventilation and air conditioning and other divisions; building water supply, drainage and heating are divided into indoor water supply system, indoor drainage system, indoor hot water system, sanitary ware, indoor heating system, outdoor water supply network and other sub-divisions; indoor water supply system is divided into indoor water supply pipes, indoor water supply pipe accessories, indoor water supply equipment, indoor fire hydrant pipes, indoor fire hydrant pipe accessories, indoor fire hydrant equipment, indoor fire sprinkler pipes and other engineering objects.

[0041] In different design specifications, different depths of delivered design drawings can be designed for each engineering object.

[0042] For example, Figure 2As shown in the figure, in a certain design specification, the depth is divided into four levels: G1, G2, G3, and G4. Among them, G1 satisfies the geometric expression of two-dimensional or symbolic recognition needs, G2 satisfies the geometric expression of rough recognition needs such as space occupancy, basic shape and overall size (area, height, volume), G3 satisfies the geometric expression of main geometric features and key dimensional information such as main dimensions and installation dimensions, G4 satisfies detailed geometric features and precise dimensions, and should show necessary detailed features and internal composition. Components should include detailed information that will be used in subsequent project stages such as construction quantity calculation, product management, manufacturing and processing.

[0043] At the level of divisions, sub-divisions and engineering objects of the construction engineering structure, the similarity of the design drawings of each engineering object under the same division is relatively high, while the similarity of the design drawings of each engineering object under different divisions is relatively low.

[0044] In the server side of the BIM platform, such as Figure 3 As shown, an encoder and multiple deep detection networks (DDN) can be constructed based on deep learning in an offline environment, and the design drawings of historical engineering objects can be used as samples to train the encoder and multiple deep detection networks DDN.

[0045] Among them, the encoder adapts to multiple divisions of engineering objects and is a general module used to extract features from design drawings.

[0046] The depth detection network is adapted to the engineering objects of a single division and is a dedicated module used to measure the depth based on the features of the design drawings.

[0047] In one embodiment of the present application, step 101 may include the following steps:

[0048] Step 1011: extract sample image data from object models of various engineering objects under multiple divisions.

[0049] In this embodiment, the division can be identified from the design drawing based on the division logo, the sub-division can be identified in the division based on the sub-division logo, and the engineering object can be identified in the sub-division based on the division item identification code, so as to read the object model obtained by three-dimensional modeling of the engineering object in the design drawing, and capture multiple frames of image data from different visual perspectives of the object model and record them as sample image data.

[0050] Step 1012: For the same engineering object, two frames of sample image data having a difference in depth are selected as sample pairs.

[0051] In different construction projects, design drawings are often constructed for the same engineering object. According to different design specifications, different delivery depths are often designed for the same engineering object in different construction projects. In addition, the boundaries of various depths are relatively fuzzy, and subjective evaluation is mainly used during the review. Therefore, for the same engineering object, any two frames of sample image data with obvious differences in depth can be screened out from the design drawings of different construction projects. As sample pairs, this can provide a reference-free depth evaluation training method for the encoder and multiple depth detection networks.

[0052] In the specific implementation, taking into account the differences in drawing habits of different designers and the different visual interception of sample image data, for the same engineering object, algorithms such as SIFT (Scale Invariant Feature Transform) can be used to calculate the similarity between two frames of sample image data, and compare the similarity with the preset similarity threshold. Since the similarity is used for preliminary screening, the similarity threshold is a lower threshold, such as 0.5.

[0053] If the similarity is greater than the preset similarity threshold, it means that the same engineering object drawn in the two frames of sample image data is highly similar. Then, Sobel, Prewitt, Canny and other algorithms can be used to detect edge data representing the depth in the two frames of sample image data respectively. The edge data can reflect the depth of the engineering object in the sample image data to a certain extent.

[0054] The difference in data quantity between edge data of two frames of sample image data is calculated as a quantity deviation value, and the quantity deviation value is compared with a preset quantity threshold value.

[0055] If the quantity deviation value is greater than the preset quantity threshold, it means that there is a relatively obvious difference in depth between the two frames of sample image data, and the two frames of sample image data are set as a sample pair.

[0056] Depth difference information is marked in the sample pair; wherein the depth difference information indicates that the depth of the edge data with the largest data volume is greater than the depth of the edge data with the smallest data volume.

[0057] Furthermore, the sample pairs may be submitted to a technician for review to ensure the accuracy of the sample pairs.

[0058] Step 1013: load the classification network and multiple deep detection networks.

[0059] In this embodiment, the classification network and multiple deep detection networks can be loaded into the memory for operation, wherein the classification network includes an encoder Encoder and a classification head structure ClassHead.

[0060] In practical applications, you can reuse third-party trained convolutional neural networks (CNNs) suitable for image processing, such as ResNet, AlexNet, FractlNet, etc., set an encoder for the feature extraction part of the structure, and replace its output layer with a classification head structure (such as one or more fully connected layers) to obtain a classification network.

[0061] Step 1014: Train a classification network based on the sample image data so that the classification network is used to divide the sample image data into multiple parts.

[0062] In this embodiment, the classification network (i.e., the encoder and the classification head structure) can be supervised trained with the sample image data as the sample and the division to which it belongs as the label, so that the classification network has the ability to divide the sample image data into multiple divisions. At this time, the encoder has the ability to extract features for dividing the multiple divisions.

[0063] In a specific implementation, the classification network can be trained through multiple rounds of iterations.

[0064] In each round of iterative training, the sample image data is input into the encoder to extract features for dividing multiple parts, which are recorded as the first sample image features.

[0065] The first sample image feature is input into the classification head structure to perform a multi-classification operation, and the probability of the sample image data belonging to multiple divisions is predicted.

[0066] The parameters of the encoder and the parameters of the classification head structure are updated according to the multi-classification loss value (such as the cross entropy loss value, etc.) between the actual division to which the sample image feature belongs and the probability that the sample image data belongs to multiple divisions.

[0067] In this process, the multi-classification loss value is substituted into optimization algorithms such as SGD (stochastic gradient descent) and Adam (Adaptive momentum) to calculate the update amplitude of the parameters in the classification network (including the parameters of the encoder and the parameters of the classification head structure), and the parameters in the classification network (including the parameters of the encoder and the parameters of the classification head structure) are updated according to the update amplitude.

[0068] In each round of iterative training, it is determined whether the preset stopping conditions are met, for example, the number of iterations reaches a certain threshold, the multi-classification loss value is less than a certain threshold, the change in the multi-classification loss value is less than a certain threshold, and so on.

[0069] If the stopping condition is met, the classification network (including the encoder and classification head structure) can be considered to have completed training. If the stopping condition is not met, the next round of iterative training can be entered.

[0070] Step 1015: If the classification network has completed training, the encoder is retained and the classification head structure is discarded.

[0071] If the classification network has completed training, the encoder can be retained and the classification head structure can be discarded. That is, the classification head structure assists in training the encoder and does not participate in the task of online detection depth.

[0072] Step 1016: With the assistance of the encoder, train the deep detection network based on the sample pairs.

[0073] In practical applications, the encoder can provide common and basic features for each deep detection network. Then, the encoder and the deep detection network can be regarded as a complete deep learning model. At this time, the deep detection network can be self-supervised trained with sample pairs as samples.

[0074] In one embodiment of the present application, step 1016 may further include the following steps:

[0075] Step 10161: The depth detection network inherits some parameters of the pre-trained edge detection network.

[0076] In this embodiment, an edge detection network whose structure matches the depth detection network can be found, such as DexiNed, LDC, etc. If the edge detection network has been pre-trained by a third party so that the edge detection network has the ability to detect image edges, then during the initialization stage of the depth detection network, the depth detection network can inherit some parameters of the edge detection network.

[0077] On the one hand, inheriting some parameters of the edge detection network can enable the deep detection network to initially possess the ability of image edge detection, complete fine-tuning, and improve training efficiency; on the other hand, the ability of image edge detection is matched with sample pairs, which facilitates the deep detection network to learn the ability of image depth detection.

[0078] For example, Figure 3 As shown, the deep detection network DDN includes a first convolution block ConvBlock_1, a second convolution block ConvBlock_2, a third convolution block ConvBlock_3, a fourth convolution block ConvBlock_4, a first pooling layer Pooling_1, a second pooling layer Pooling_2, a third pooling layer Pooling_3 and a fully connected layer FC.

[0079] Among them, the first convolution block ConvBlock_1, the second convolution block ConvBlock_2, the third convolution block ConvBlock_3, and the fourth convolution block ConvBlock_4 are all block structures related to convolution, for example, multiple convolutional layers (ConvolutionalLayer), BN (BatchNormlization), ReLU (Rectified Linear Unit, linear rectification function), and so on.

[0080] In this example, the parameters of the first convolution block ConvBlock_1, the second convolution block ConvBlock_2, the third convolution block ConvBlock_3, the fourth convolution block ConvBlock_4, the first pooling layer Pooling_1, the second pooling layer Pooling_2, and the third pooling layer Pooling_3 are all inherited from the edge detection network.

[0081] In addition, the parameters of the fully connected layer FC can be randomly set.

[0082] Step 10162: If the inheritance is completed, the two frames of sample image data in the sample pair are respectively input into the encoder to extract the second sample image features used for dividing the multiple parts.

[0083] When the deep detection network DDN is initialized, the two frames of sample image data in the sample pair are respectively input into the encoder Encoder, and the encoder Encoder extracts features for dividing multiple parts from the two frames of sample image data, which are recorded as second sample image features.

[0084] Step 10163: For the same division, input the second sample image feature into the depth detection network to identify the sample depth value of the corresponding object model.

[0085] The encoder outputs the second sample image feature to the depth detection network DDN corresponding to the division to which the sample image data belongs. The depth detection network DDN processes the second sample image feature according to its structure, thereby outputting the depth of the object model corresponding to the sample image data, which is recorded as the sample depth value.

[0086] For example, Figure 3 As shown, for the same part, the second sample image feature is input into the first convolution block ConvBlock_1 to perform a convolution operation to obtain the first sample edge feature.

[0087] The first sample edge feature is input into the first pooling layer Pooling_1 to perform a maximum pooling operation to obtain the second sample edge feature.

[0088] The second sample edge feature is input into the second convolution block ConvBlock_2 to perform a convolution operation to obtain the third sample edge feature.

[0089] The third sample edge feature is input into the second pooling layer Pooling_2 to perform a maximum pooling operation to obtain a fourth sample edge feature.

[0090] The fourth sample edge feature is input into the third convolution block ConvBlock_3 to perform a convolution operation to obtain the fifth sample edge feature.

[0091] Input the fifth sample edge feature into the third pooling layer Pooling_3 to perform a maximum pooling operation to obtain a sixth sample edge feature;

[0092] Use Concat and other functions to concatenate the second sample edge feature, the fourth sample edge feature and the sixth sample edge feature into the seventh sample edge feature, thereby fusing features at multiple levels together, increasing the amount of feature information and improving the accuracy of depth detection.

[0093] The seventh sample edge feature is input into the fourth convolution block ConvBlock_4 to perform a convolution operation to obtain the eighth sample edge feature.

[0094] The eighth sample edge feature is input into the fully connected layer FC and mapped into the sample depth value of the object model.

[0095] Step 10164: Under the condition of maintaining the encoder parameters unchanged, update the parameters of the depth detection network according to the pairing loss value between the depth difference information and the sample depth value.

[0096] In a specific implementation, the pairwise loss value (Pairwise Loss) between the depth difference information and the sample depth value can be calculated, for example, RankNet Loss (ranking loss), FRNet Loss (feature refinement loss), MarginLoss (margin loss), and so on.

[0097] Substitute the pairing loss value into optimization algorithms such as SGD and Adam, calculate the update amplitude of the parameters in the deep learning model (including the parameters of the encoder and the parameters of the deep detection network), and update the parameters in the deep learning model (including the parameters of the encoder and the parameters of the deep detection network) according to the update amplitude. When the update of the parameters of the deep detection network is completed, the encoder parameters are not updated, and the encoder parameters are kept unchanged.

[0098] In each round of iterative training, it is determined whether the preset stopping conditions are met, for example, the number of iterations reaches a certain threshold, the multi-classification loss value is less than a certain threshold, the change in the pairing loss value is less than a certain threshold, and so on.

[0099] If the stopping condition is met, the deep learning model (including the encoder and the deep detection network) can be considered to have completed training. If the stopping condition is not met, the next round of iterative training can be entered.

[0100] Step 1017: If the depth detection network has completed training, the encoder is cascaded with the depth detection network.

[0101] When each deep detection network has completed training, each deep detection network can be cascaded after the encoder to facilitate the execution of the deep detection task.

[0102] Step 102: When the construction project is digitally delivered, a three-dimensional design drawing of the building is received.

[0103] When designers digitally deliver construction projects, they can receive three-dimensional design drawings of designated buildings. The design drawings can be converted into GDB, IFC, RVT, SU and other formats to meet the requirements of data transmission, drawing review, CIM (City Information Modeling) and other requirements.

[0104] Step 103: In the design drawing, target image data is captured for the object models of various engineering objects under multiple divisions.

[0105] In this embodiment, the division can be identified from the design drawing based on the division mark, the sub-division can be identified in the division based on the sub-division mark, and the engineering object can be identified in the sub-division based on the division item identification code, so as to read the object model obtained by three-dimensional modeling of the engineering object in the design drawing, and capture multiple frames of image data from different visual perspectives of the object model and record them as target image data.

[0106] Step 104: Input the target image data into the encoder to extract common image features for dividing the multiple parts.

[0107] In this embodiment, if Figure 3 As shown, each frame of target image data can be input into the encoder. The encoder extracts features for dividing the target image data into multiple parts based on its training capabilities and records them as common image features.

[0108] Step 105: For the same division, the common image features are input into the depth detection network to identify the target depth value of the object model.

[0109] In this embodiment, if Figure 3As shown, the encoder outputs the common image features to the depth detection network DDN corresponding to the division to which the target image data belongs. The depth detection network DDN processes the common image features according to its structure, thereby outputting the depth of the object model corresponding to the target image data, which is recorded as the target depth value.

[0110] For example, Figure 3 As shown, for the same division, the common image features are input into the first convolution block ConvBlock_1 to perform a convolution operation to obtain the first target edge features.

[0111] The first target edge feature is input into the first pooling layer Pooling_1 to perform a maximum pooling operation to obtain the second target edge feature.

[0112] The second target edge feature is input into the second convolution block ConvBlock_2 to perform a convolution operation to obtain the third target edge feature.

[0113] The third target edge feature is input into the second pooling layer Pooling_2 to perform a maximum pooling operation to obtain a fourth target edge feature.

[0114] The fourth target edge feature is input into the third convolution block ConvBlock_3 to perform a convolution operation to obtain the fifth target edge feature.

[0115] The fifth target edge feature is input into the third pooling layer Pooling_3 to perform a maximum pooling operation to obtain a sixth target edge feature.

[0116] Use functions such as Concat to concatenate the second target edge feature, the fourth target edge feature, and the sixth target edge feature into the seventh target edge feature, thereby fusing features at multiple levels together, increasing the amount of feature information, and improving the accuracy of depth detection.

[0117] The seventh target edge feature is input into the fourth convolution block ConvBlock_4 to perform a convolution operation to obtain the eighth target edge feature.

[0118] The eighth target edge feature is input into the fully connected layer FC map to the target depth value of the object model.

[0119] Step 106: classify the object model in depth according to the target depth value.

[0120] In this embodiment, the object model can be graded in depth according to the target depth value of the object model, wherein the grade is positively correlated with the target depth value, that is, the larger the target depth value, the higher the grade, and vice versa, the smaller the target depth value, the lower the grade.

[0121] In a specific implementation, a preset depth grading table is loaded; wherein the depth grading table records the mapping relationship between multiple depth ranges and levels.

[0122] The target depth value is compared with each depth range, so that the level corresponding to the depth range in which the target depth value is located is set as the depth level of the object model.

[0123] Step 107: Generate a quality value of depth for the design drawing according to the grade.

[0124] In practical applications, the quality value of the design drawings can be evaluated based on the level of each engineering object in depth.

[0125] In one embodiment of the present application, step 107 may include the following steps:

[0126] Step 1071: Query the stage of digital delivery of the engineering object.

[0127] On the BIM server side, you can query the stage of digital delivery of the project object, such as the preliminary design stage, in-depth design stage, etc.

[0128] Step 1072: Query the delivery depth and corresponding constraints set for the project object at the stage.

[0129] At different stages, the delivery depth (level) and constraints can be set for the same project object.

[0130] For example, in the preliminary design stage, the delivery depth of the object models of most engineering objects is G2 and N2, and the delivery depth of the object models of a small number of engineering objects (such as the main building, structural main body, water supply and drainage main network, main electrical bridge, HVAC main ventilation duct, etc.) is G3 and N3; in the in-depth design stage, the delivery depth of the object models of all engineering objects is G4 and N4.

[0131] In addition, constraints represent the selectivity of the delivery depth, which includes mandatory and optional.

[0132] For example, in building water supply, drainage and heating, the delivery depth of all engineering objects in the indoor water supply system is G3 and N4, which is mandatory; in building roofing, the delivery depth of all engineering objects in insulation and heat insulation (sub-division) (such as plate material insulation layer, fiber material insulation layer, etc.) is N4, which is optional, and so on.

[0133] Step 1073: Subtract the delivery depth from the level to obtain a depth deviation value.

[0134] When the depth grade and the delivery depth of the engineering object are dimensionless, the depth deviation value may be obtained by subtracting the corresponding delivery depth from the depth grade of the engineering object. The depth deviation value may be a positive value or an amplitude value.

[0135] For example, the level of the engineering object in depth is G3, the delivered depth is G4, and the depth deviation value is -1; the level of the engineering object in depth is G3, the delivered depth is G2, and the depth deviation value is 1.

[0136] Step 1074: merge the depth deviation value into the quality value of the design drawing with respect to depth according to the constraint condition.

[0137] In this embodiment, the depth deviation value can be integrated into the quality value of the design drawing with respect to the depth in a linear or nonlinear manner based on the constraint condition.

[0138] Exemplarily, weights are configured for corresponding depth deviation values ​​according to constraints; wherein, when the constraints are mandatory, the weight is a first value, and when the constraints are optional, the weight is a second value; and the first value is greater than the second value.

[0139] The product of each depth deviation value and the weight is summed to obtain the quality value of the design drawing with respect to depth.

[0140] In this example, the quality value of the design drawing with respect to depth can be expressed as:

[0141]

[0142] Among them, i is the number of the engineering object, i∈N, N is the set of object models drawn for the engineering object in the same design drawing, Score is the quality value of the design drawing with respect to depth, DeepDel i It is the constraint condition set for the delivery depth of the engineering object. M is mandatory and O is optional. DeepDev i is the depth deviation value of the object model drawn for the engineering object, w i is the weight, αα is the first value, β is the second value, αα>β.

[0143] Step 108: Generate review information on digital delivery for the design drawings based on the quality value.

[0144] In this embodiment, the quality value of the depth of the design drawings can be written into the review information of the design drawings regarding digital delivery, and a simple evaluation can be performed using thresholds, classification levels, etc., which can be used as one of the automated review information to provide a reference for self-review, preliminary review, and review.

[0145] In this embodiment, encoders adapted to multiple divisions and multiple depth detection networks adapted to a single division are trained; when digital delivery of a construction project is performed, a three-dimensional design drawing of a building is received; in the design drawing, target image data is intercepted for the object model of various engineering objects under multiple divisions; the target image data is input into the encoder to extract the common image features used to divide the multiple divisions; for the same division, the common image features are input into the depth detection network to identify the target depth value of the object model; the object model is graded in depth according to the target depth value; a quality value about the depth is generated for the design drawing according to the grade; and audit information about digital delivery is generated for the design drawing according to the quality value. This embodiment provides automated depth auditing in the building information model platform, provides reference for self-examination, preliminary examination, re-examination and other auditing processes, reduces the reliance on manual auditing of design drawings, greatly reduces labor costs, and, based on deep learning to determine the level of depth, it can reduce the influence of subjectivity of manual auditing, provide a standardized scale, and improve the accuracy of the level of depth, thereby effectively improving the efficiency of auditing the depth of object models in design drawings.

[0146] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0147] Reference Figure 4 , shows a schematic diagram of a design drawing review device provided in an embodiment of the present application, which is applied to a building information modeling platform. The device may specifically include the following modules:

[0148] Model training module 401, used to train encoders adapted to multiple divisions and multiple deep detection networks adapted to a single division;

[0149] The design drawing receiving module 402 is used to receive the three-dimensional design drawings of the building when the construction project is digitally delivered;

[0150] The target image data interception module 403 is used to intercept the target image data of the object models of various engineering objects under the multiple divisions in the design drawings;

[0151] A common image feature extraction module 404, used for inputting the target image data into the encoder to extract common image features for dividing the multiple sub-parts;

[0152] A target depth value recognition module 405 is used for inputting the common image features into the depth detection network to recognize a target depth value of the object model for the same sub-part;

[0153] A depth level division module 406, configured to divide the object model into levels in depth according to the target depth value;

[0154] A quality value generating module 407, configured to generate a quality value of depth for the design drawing according to the grade;

[0155] The audit information generation module 408 is used to generate audit information about digital delivery for the design drawings based on the quality value.

[0156] Optionally, the model training module 401 includes:

[0157] A sample image data interception module is used to intercept sample image data from object models of various engineering objects under multiple divisions;

[0158] A sample pair screening module, for screening out two frames of sample image data having a difference in depth as a sample pair for the same engineering object;

[0159] A network loading module, used to load a classification network and multiple deep detection networks; the classification network includes an encoder and a classification head structure;

[0160] A classification network training module, used for training the classification network according to the sample image data, so that the classification network is used to divide the sample image data into a plurality of sub-parts;

[0161] An encoder retention module, used for retaining the encoder and discarding the classification head structure if the classification network completes training;

[0162] A deep detection network training module, used for training the deep detection network according to the sample pairs with the assistance of the encoder;

[0163] A network cascade module is used to cascade the encoder with the deep detection network if the deep detection network has completed training.

[0164] Optionally, the sample pair screening module is further used for:

[0165] For the same engineering object, calculating the similarity between two frames of sample image data;

[0166] If the similarity is greater than a preset similarity threshold, detecting edge data representing depth in the sample image data;

[0167] Calculating the difference in data quantity between the edge data of the two frames of the sample image data as a quantity deviation value;

[0168] If the quantity deviation value is greater than a preset quantity threshold, setting the two frames of sample image data as a sample pair;

[0169] Depth difference information is marked in the sample pair; the depth difference information indicates that the depth of the edge data with the largest data volume is greater than the depth of the edge data with the smallest data volume.

[0170] Optionally, the classification network training module is further used for:

[0171] Inputting the sample image data into the encoder to extract a first sample image feature for dividing the plurality of sub-parts;

[0172] Inputting the first sample image feature into the classification head structure to predict the probability that the sample image data belongs to multiple divisions;

[0173] Update the parameters of the encoder and the parameters of the classification head structure according to the multi-classification loss value between the division to which the sample image feature actually belongs and the probability that the sample image data belongs to multiple divisions;

[0174] The depth detection network training module includes:

[0175] A network initialization module, used to inherit some parameters of the pre-trained edge detection network for the deep detection network;

[0176] A sample image feature extraction module, for inputting the sample image data of two frames in the sample pair into the encoder to extract second sample image features for dividing the plurality of sub-sections if the inheritance is completed;

[0177] A sample depth value recognition module, for inputting the second sample image feature into the depth detection network to recognize a sample depth value of the object model for the same sub-portion;

[0178] A network parameter updating module is used to update the parameters of the depth detection network according to the pairing loss value between the depth difference information and the sample depth value under the condition of maintaining the parameters of the encoder unchanged.

[0179] Optionally, the depth detection network includes a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a first pooling layer, a second pooling layer, a third pooling layer and a fully connected layer; the parameters of the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the first pooling layer, the second pooling layer and the third pooling layer are all inherited from the edge detection network;

[0180] The sample depth value recognition module is also used for:

[0181] For the same said sub-portion, inputting the second sample image feature into the first convolution block to perform a convolution operation, so as to obtain a first sample edge feature;

[0182] Inputting the first sample edge feature into the first pooling layer to perform a maximum pooling operation to obtain a second sample edge feature;

[0183] Inputting the second sample edge feature into the second convolution block to perform a convolution operation to obtain a third sample edge feature;

[0184] Inputting the third sample edge feature into the second pooling layer to perform a maximum pooling operation to obtain a fourth sample edge feature;

[0185] Inputting the fourth sample edge feature into the third convolution block to perform a convolution operation to obtain a fifth sample edge feature;

[0186] Inputting the fifth sample edge feature into the third pooling layer to perform a maximum pooling operation to obtain a sixth sample edge feature;

[0187] splicing the second sample edge feature, the fourth sample edge feature and the sixth sample edge feature into a seventh sample edge feature;

[0188] Inputting the seventh sample edge feature into the fourth convolution block to perform a convolution operation to obtain an eighth sample edge feature;

[0189] The eighth sample edge feature is input into the fully connected layer and mapped into a sample depth value of the object model.

[0190] Optionally, the target depth value recognition module 405 is further used for:

[0191] For the same sub-portion, inputting the common image feature into the first convolution block to perform a convolution operation to obtain a first target edge feature;

[0192] Inputting the first target edge feature into the first pooling layer to perform a maximum pooling operation to obtain a second target edge feature;

[0193] Inputting the second target edge feature into the second convolution block to perform a convolution operation to obtain a third target edge feature;

[0194] Inputting the third target edge feature into the second pooling layer to perform a maximum pooling operation to obtain a fourth target edge feature;

[0195] Inputting the fourth target edge feature into the third convolution block to perform a convolution operation to obtain a fifth target edge feature;

[0196] Inputting the fifth target edge feature into the third pooling layer to perform a maximum pooling operation to obtain a sixth target edge feature;

[0197] splicing the second target edge feature, the fourth target edge feature and the sixth target edge feature into a seventh target edge feature;

[0198] Inputting the seventh target edge feature into the fourth convolution block to perform a convolution operation to obtain an eighth target edge feature;

[0199] The eighth target edge feature is input into the fully connected layer and mapped into a target depth value of the object model.

[0200] Optionally, the depth level division module 406 is further configured to:

[0201] Loading a preset depth classification table; the depth classification table records the mapping relationship between multiple depth ranges and levels;

[0202] Setting the level corresponding to the depth range of the target depth value as the depth level of the object model;

[0203] The quality value generating module 407 includes:

[0204] A stage query module, used to query the stage of digital delivery of the engineering object;

[0205] A delivery requirement query module, used to query the delivery depth and constraint conditions set for the engineering object at the stage;

[0206] A depth deviation value calculation module, used for subtracting the delivery depth from the level to obtain a depth deviation value;

[0207] A depth deviation value fusion module is used to fuse the depth deviation value into the quality value of the design drawing with respect to depth according to the constraint condition.

[0208] Optionally, the depth deviation value fusion module is further used for:

[0209] According to the constraint condition, a weight is configured for the depth deviation value; wherein when the constraint condition is mandatory, the weight is a first value, and when the constraint condition is optional, the weight is a second value; and the first value is greater than the second value;

[0210] The products of the depth deviation values ​​and the weights are summed to obtain the quality value of the design drawing with respect to depth.

[0211] An embodiment of the present application provides a design drawing review device, and by using the device, each step in the aforementioned method embodiments can be implemented.

[0212] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment part.

[0213] Reference Figure 5 , shows a schematic diagram of a terminal device provided by an embodiment of the present application. Figure 5 As shown, the terminal device 400 in the embodiment of the present application includes: a processor 410, a memory 420, and a computer program 421 stored in the memory 420 and executable on the processor 410. When the processor 410 executes the computer program 421, the steps in each embodiment of the above-mentioned design drawing review method are implemented. Alternatively, when the processor 410 executes the computer program 421, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0214] Exemplarily, the computer program 421 may be divided into one or more modules / units, which are stored in the memory 420 and executed by the processor 410 to complete the present application. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which may be used to describe the execution process of the computer program 421 in the terminal device 400.

[0215] The terminal device 400 may include, but is not limited to, a processor 410 and a memory 420. Those skilled in the art will appreciate that Figure 5 It is only an example of the terminal device 400 and does not constitute a limitation on the terminal device 400. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device 400 may also include input and output devices, network access devices, buses, etc.

[0216] The processor 410 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0217] The memory 420 may be an internal storage unit of the terminal device 400, such as a hard disk or memory of the terminal device 400. The memory 420 may also be an external storage device of the terminal device 400, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 400. Further, the memory 420 may also include both an internal storage unit of the terminal device 400 and an external storage device. The memory 420 is used to store the computer program 421 and other programs and data required by the terminal device 400. The memory 420 may also be used to temporarily store data that has been output or is to be output.

[0218] An embodiment of the present application also discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for reviewing design drawings as described in the aforementioned embodiments is implemented.

[0219] The embodiments of the present application further disclose a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the design drawing review method as described in the aforementioned embodiments is implemented.

[0220] The embodiments of the present application further disclose a computer program product. When the computer program product is run on a computer, the computer is enabled to execute the design drawing review method described in the aforementioned embodiments.

[0221] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application is described in detail with reference to the above-mentioned embodiments, a person skilled in the art should understand that the technical solutions described in the above-mentioned embodiments can still be modified, or some of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for reviewing design drawings, characterized in that: Applied to a building information modeling platform, the method comprises: Training encoders adapted to multiple parts and multiple deep detection networks adapted to a single part; Receive constructed, three-dimensional design drawings of the building when digitally delivering the construction project; In the design drawing, target image data is captured for object models of various engineering objects under a plurality of divisions; Inputting the target image data into the encoder to extract common image features for dividing a plurality of sub-parts; For the same sub-portion, inputting the common image features into the depth detection network to identify a target depth value of the object model; classifying the object model in depth according to the target depth value; generating a quality value of depth for the design drawing according to the grade; Audit information on digital delivery is generated for the design drawings according to the quality value.

2. The method according to claim 1, characterized in that The training of encoders adapted to multiple divisions and multiple deep detection networks adapted to a single division includes: Extract sample image data from object models of various engineering objects under multiple divisions; For the same engineering object, two frames of sample image data having a difference in depth are selected as sample pairs; Loading a classification network and multiple deep detection networks; the classification network includes an encoder and a classification head structure; Training the classification network according to the sample image data so that the classification network is used to divide the sample image data into a plurality of parts; If the classification network completes training, retaining the encoder and discarding the classification head structure; With the assistance of the encoder, training the deep detection network based on the sample pairs; If the deep detection network has completed training, the encoder is cascaded with the deep detection network.

3. The method according to claim 2, characterized in that For the same engineering object, screening out two frames of sample image data having a difference in depth as a sample pair includes: For the same engineering object, calculating the similarity between two frames of sample image data; If the similarity is greater than a preset similarity threshold, detecting edge data representing depth in the sample image data; Calculating the difference in data quantity between the edge data of the two frames of the sample image data as a quantity deviation value; If the quantity deviation value is greater than a preset quantity threshold, setting the two frames of sample image data as a sample pair; Depth difference information is marked in the sample pair; the depth difference information indicates that the depth of the edge data with the largest data volume is greater than the depth of the edge data with the smallest data volume.

4. The method according to claim 3, characterized in that The step of training the classification network according to the sample image data so that the classification network is used to divide the sample image data into a plurality of parts comprises: Inputting the sample image data into the encoder to extract a first sample image feature for dividing the plurality of sub-parts; Inputting the first sample image feature into the classification head structure to predict the probability that the sample image data belongs to multiple divisions; Update the parameters of the encoder and the parameters of the classification head structure according to the multi-classification loss value between the division to which the sample image feature actually belongs and the probability that the sample image data belongs to multiple divisions; The step of training the depth detection network according to the sample pairs with the assistance of the encoder includes: Inheriting some parameters of the pre-trained edge detection network to the depth detection network; If the inheritance is completed, the sample image data of the two frames in the sample pair are respectively input into the encoder to extract the second sample image features used for dividing the multiple sub-parts; For the same sub-portion, inputting the second sample image feature into the depth detection network to identify the sample depth value of the object model; Under the condition of keeping the parameters of the encoder unchanged, the parameters of the depth detection network are updated according to the pairing loss value between the depth difference information and the sample depth value.

5. The method according to claim 4, characterized in that The depth detection network includes a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a first pooling layer, a second pooling layer, a third pooling layer and a fully connected layer; the parameters of the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the first pooling layer, the second pooling layer and the third pooling layer are all inherited from the edge detection network; For the same sub-portion, inputting the second sample image feature into the depth detection network to identify the sample depth value of the object model includes: For the same said sub-portion, inputting the second sample image feature into the first convolution block to perform a convolution operation, so as to obtain a first sample edge feature; Inputting the first sample edge feature into the first pooling layer to perform a maximum pooling operation to obtain a second sample edge feature; Inputting the second sample edge feature into the second convolution block to perform a convolution operation to obtain a third sample edge feature; Inputting the third sample edge feature into the second pooling layer to perform a maximum pooling operation to obtain a fourth sample edge feature; Inputting the fourth sample edge feature into the third convolution block to perform a convolution operation to obtain a fifth sample edge feature; Inputting the fifth sample edge feature into the third pooling layer to perform a maximum pooling operation to obtain a sixth sample edge feature; splicing the second sample edge feature, the fourth sample edge feature and the sixth sample edge feature into a seventh sample edge feature; Inputting the seventh sample edge feature into the fourth convolution block to perform a convolution operation to obtain an eighth sample edge feature; The eighth sample edge feature is input into the fully connected layer and mapped into a sample depth value of the object model.

6. The method according to claim 5, characterized in that For the same sub-portion, inputting the common image features into the depth detection network to identify the target depth value of the object model includes: For the same sub-portion, inputting the common image feature into the first convolution block to perform a convolution operation to obtain a first target edge feature; Inputting the first target edge feature into the first pooling layer to perform a maximum pooling operation to obtain a second target edge feature; Inputting the second target edge feature into the second convolution block to perform a convolution operation to obtain a third target edge feature; Inputting the third target edge feature into the second pooling layer to perform a maximum pooling operation to obtain a fourth target edge feature; Inputting the fourth target edge feature into the third convolution block to perform a convolution operation to obtain a fifth target edge feature; Inputting the fifth target edge feature into the third pooling layer to perform a maximum pooling operation to obtain a sixth target edge feature; splicing the second target edge feature, the fourth target edge feature and the sixth target edge feature into a seventh target edge feature; Inputting the seventh target edge feature into the fourth convolution block to perform a convolution operation to obtain an eighth target edge feature; The eighth target edge feature is input into the fully connected layer and mapped into a target depth value of the object model.

7. The method according to any one of claims 1 to 6, characterized in that The step of classifying the object model in depth according to the target depth value comprises: Loading a preset depth classification table; the depth classification table records the mapping relationship between multiple depth ranges and levels; Setting the level corresponding to the depth range of the target depth value as the depth level of the object model; Generating a quality value of depth for the design drawing according to the grade includes: Query the stage of digital delivery of the engineering object; Query the delivery depth and constraints set for the engineering object at the stage; Subtract the delivery depth from the level to obtain a depth deviation value; The depth deviation value is merged into the quality value of the design drawing with respect to the depth according to the constraint condition.

8. The method according to claim 7, characterized in that The step of fusing the depth deviation value into a quality value of the design drawing with respect to depth according to the constraint condition comprises: According to the constraint condition, a weight is configured for the depth deviation value; wherein when the constraint condition is mandatory, the weight is a first value, and when the constraint condition is optional, the weight is a second value; and the first value is greater than the second value; The products of the depth deviation values ​​and the weights are summed to obtain the quality value of the design drawing with respect to depth.

9. A terminal 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 computer program, the method for reviewing design drawings as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for reviewing design drawings as described in any one of claims 1 to 8 is implemented.

Images