Engineering drawing steel bar bulk sample annotation text detection method, system and equipment based on deep learning and storage medium

Abstract

The invention provides an engineering drawing steel bar bulk sample annotation text detection method, system and device based on deep learning, and a storage medium. The method comprises the following steps: S1, preprocessing an engineering drawing image and labeling a steel bar bulk sample annotation text region data set; s2, constructing a staged multi-scale feature map extraction backbone network framework; s3, introducing compression excitation into each scale feature map; s4, adopting an ACON adaptive activation function in nonlinear modeling; s5, constructing a multi-scale feature pyramid structure based on the RSE-FPN, and superposing scales to reinforce fusion; and S6, outputting the approximate binary image to realize accurate prediction of the bounding box. According to the method, the OCR detection precision is remarkably improved, small character and complex background interference is effectively overcome, and missing detection and false detection are reduced; the model is lightweight to facilitate efficient deployment of equipment; the output boundary is compact and accurate, the semantics is reasonable, and subsequent recognition and analysis are facilitated; the method has excellent generalization ability, adapts to steel bar large sample drawings of different formats, definitions and styles, and is high in engineering practicability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and engineering drawing processing, specifically relating to a method, system, device and storage medium for detecting reinforcement detail annotation text in engineering drawings based on deep learning. Background Technology

[0002] In the field of engineering design and construction, structural drawings serve as a crucial technical basis for projects, and their accuracy directly impacts project quality and safety. Among these, the reinforcement detail drawing is a vital component, used to describe the dimensions, numbering, and specifications of reinforced concrete members. Reinforcement detail drawings often contain explanatory text indicating the number, diameter, length, and number of reinforcing bars, providing essential information for construction personnel in reinforcement selection and fabrication.

[0003] Traditional drawing analysis relies on manual review, which is not only time-consuming and labor-intensive but also prone to oversights. In recent years, with the rapid development of image recognition technology, automated drawing interpretation has gradually become a research hotspot. However, due to the characteristics of engineering drawings, such as complex backgrounds, overlapping graphic elements, dense text arrangement, and large size spans, traditional OCR methods, such as the drawing information extraction method and device based on optical character recognition disclosed in Chinese patent CN116884028A, have poor performance in the task of recognizing reinforcement detail annotations, exhibiting problems such as high false negative rate, high false recognition rate, poor generalization, and weak anti-interference ability.

[0004] Current mainstream text detection methods primarily rely on deep learning frameworks such as DBNet (Differentiable Binarization Network) and U-Net for object detection in images. For example, Chinese patent CN117975464A discloses a method and system for recognizing text information in electrical secondary drawings based on U-Net. While this invention can achieve end-to-end text region localization, it still faces significant challenges when dealing with complex scenarios such as occluded text, single-character annotations, non-standard fonts, and uneven text spacing in engineering drawings. Furthermore, as model complexity increases, traditional networks often struggle to operate efficiently on low-power devices, limiting their deployment capabilities on mobile and edge devices. Summary of the Invention

[0005] The first objective of this invention is to provide a deep learning-based method for detecting the annotation text of steel reinforcement details in engineering drawings, addressing the aforementioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A deep learning-based method for detecting text in reinforcement detail annotations on engineering drawings includes the following steps:

[0008] S1. Preprocessing of engineering drawing images and annotation of text regions for steel reinforcement detail annotations;

[0009] S2. Construct a multi-scale feature map extraction backbone network architecture in stages;

[0010] S3. Introduce compression excitation in the feature maps at each scale;

[0011] S4. The ACON adaptive activation function is used for nonlinear modeling.

[0012] S5. Construct a multi-scale feature pyramid structure based on RSE-FPN and superimpose scale-enhanced fusion;

[0013] S6. Output an approximate binary map to achieve accurate prediction of the bounding box.

[0014] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0015] As a preferred technical solution of the present invention: in step S1, the preprocessing of the engineering drawing image includes image format conversion, using the Sobel operator or Canny algorithm to enhance edge information, and using median filtering or bilateral filtering to remove noise interference in the image background.

[0016] As a preferred technical solution of the present invention: In step S1, the annotation of the rebar detail annotation text region dataset includes manual semantic segmentation annotation of the detail annotation region, dividing the annotation region into three categories: number region, root number diameter region and length region, and assigning region labels to the annotation logical relationship of each type of region.

[0017] As a preferred technical solution of the present invention: in step S3, compression is performed by global average pooling to condense the information of the entire feature map, and excitation is performed by two fully connected layers to learn the correlation between feature channels.

[0018] As a preferred technical solution of the present invention: in step S4, a switching factor is introduced. To learn the parameter switching between nonlinear and linear activation, for a maximum function By using the softmax approach for smooth approximation, we can obtain an approximate formula for the maximum value function:

[0019]

[0020] When n=2, the maximum value function It can be approximated as:

[0021]

[0022] Substituting the Sigmoid function into the above approximation formula, it can be expressed as:

[0023]

[0024] because and Both are linear functions, and we can let Where a and b are linear control coefficients that control the upper and lower bounds of the activation function and the activation result, we can obtain the general form of the ACON activation function:

[0025]

[0026] In the formula, Used to control whether a function is activated or not.

[0027] As a preferred technical solution of the present invention: In step S5, the RSE-FPN network is a combination of Res2Net with multi-branch receptive field enhancement and Feature Pyramid Network (FPN). The features after being concatenated at the network neck are divided into three branches: the first branch performs only one transposed convolution to upsample the features and performs scale restoration through 3×3 convolution to preserve the detection capability of the original text; the second branch performs two feature enhancement operations and aligns the dimensions with the original features through pointwise convolution; the third branch expands the receptive field through two convolution operations and aligns the feature dimensions with the pointwise convolution through upsampling; finally, the feature maps of the three branches are superimposed to achieve scale enhancement fusion for text of different areas in large sample annotations.

[0028] As a preferred embodiment of the present invention: In step S6, each pixel value in the probability map P is transformed by a differentiable activation function. To ensure that the entire process is differentiable, the Sigmoid function is used as the activation function.

[0029]

[0030] In the formula, This represents the probability value of the input. This represents the result after the Sigmoid function transformation, with values ​​ranging from (0,1), where k is the magnification factor; the other branch performs binarized threshold map prediction, and the threshold map is calculated using the following formula:

[0031]

[0032] In the formula, t is a predefined threshold, and (i,j) are the coordinate points in the graph;

[0033] The probability map after Sigmoid transformation is multiplied pixel by pixel with the threshold map to obtain an approximate threshold map;

[0034] After obtaining the probability map, threshold map, and approximate binary map, cross-entropy loss is used as the loss function.

[0035] Probabilistic graphical cross-entropy loss function The calculation formula is as follows:

[0036]

[0037] In the formula, 𝑁 represents the number of samples, and 𝐶 represents the number of categories. It is the One-Hot encoding of the real label. This represents the probability value at the corresponding position in the predicted probability map;

[0038] Threshold graph cross-entropy loss function The calculation formula is as follows:

[0039]

[0040] In the formula, This represents the threshold at the corresponding position in the predicted threshold map;

[0041] Approximate binary graph loss function The calculation formula is as follows:

[0042]

[0043] In the formula, It is the pixel value at the corresponding position in the approximate binary image;

[0044] Finally, the total loss function is obtained by weighted summation of the probabilistic graphical loss, threshold graphical loss, and approximate binarization loss. The formula is shown below:

[0045]

[0046] In the formula, Let the probabilistic graphical cross-entropy loss function be used. The threshold graph cross-entropy loss function is used. This is the loss function for an approximate binary graph.

[0047] The second objective of this invention is to provide a deep learning-based system for detecting text in engineering drawing reinforcement detail annotations, comprising the following modules:

[0048] - Image preprocessing and dataset annotation module, used for preprocessing engineering drawing images and annotating the text regions of steel bar detail annotations;

[0049] - Backbone network architecture building module, used to build a lightweight convolutional neural network based on depthwise separable convolutions as the backbone network for text feature extraction;

[0050] - The compression stimulus module is used to automatically calculate the importance weight of each channel through global average pooling and inter-channel weight recalibration mechanism, enhance the channel response containing annotated text features, and suppress irrelevant background information;

[0051] - Nonlinear modeling module, used to introduce ACON adaptive activation function, to give neurons the ability to learn whether to activate in the nonlinear transformation stage, to improve the network’s ability to distinguish complex backgrounds or blurry characters in drawings. ACON can self-adjust activation state during training, effectively improving the network’s expressive ability and adaptability.

[0052] - The pyramid structure construction and scale enhancement fusion module is used to improve the network's ability to distinguish text regions of different sizes and enhance the model's perception and expression of small characters by performing channel alignment, scale normalization and pixel-wise weighted fusion operations on feature maps at multiple scales.

[0053] - The text detection and prediction module is used to output an approximate binary map based on the obtained probability map and threshold map, and finally achieve accurate prediction of the bounding box.

[0054] A third objective of this invention is to provide an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus.

[0055] The memory is used to store computer programs;

[0056] A processor is configured to execute a computer program stored in a memory to implement the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings, as described above.

[0057] Another objective of this invention is to provide a non-volatile storage medium storing an executable program, which, when executed by a processor, implements the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings as described above.

[0058] This invention provides a method, system, device, and storage medium for detecting reinforcement detail annotation text in engineering drawings based on deep learning, which has the following beneficial effects:

[0059] 1) Significantly improved detection accuracy: By utilizing structural priors and region segmentation strategies, interference from non-target regions is effectively avoided. Combined with the SE attention mechanism and ACON adaptive activation function, the network's ability to extract key character features is enhanced, resulting in a significant improvement in detection accuracy compared to traditional OCR methods, especially in numbered regions and densely sized regions.

[0060] 2) Stronger robustness to small characters and complex backgrounds: By constructing scale-enhanced fusion and introducing the RSE-FPN structure, the model can maintain stable detection capability for small-sized, low-contrast characters while keeping it lightweight, greatly reducing missed detections and false detections. It is particularly suitable for high-density, small-character annotation situations in actual drawings.

[0061] 3) Excellent lightweight deployment performance: The overall model adopts MobileNet as the basic framework, combined with depthwise separable convolution and channel compression strategies, which significantly reduces the number of model parameters and computational complexity while ensuring detection performance. It can be efficiently deployed on mobile devices or edge computing devices to meet the needs of rapid response and real-time recognition in engineering sites.

[0062] 4) Tight boundaries and reasonable semantics in the detection results: By introducing differentiable binarization and circular structure interference removal mechanism, the text detection box output by this invention not only has precise and tight boundaries, but also combines regional semantic labels to give initial structural information, which helps subsequent text recognition and structure analysis tasks.

[0063] 5) Possesses excellent engineering adaptability and generalization ability: It is applicable to steel reinforcement detail drawings of different formats, resolutions and drawing styles, and has good versatility and potential for engineering promotion. Attached Figure Description

[0064] Figure 1 The flowchart shows the deep learning-based method for detecting the annotation text of steel reinforcement details in engineering drawings provided by this invention.

[0065] Figure 2 This is a network architecture diagram for text detection in engineering drawings.

[0066] Figure 3 This is a schematic diagram for extracting text features from engineering drawings.

[0067] Figure 4 This is a schematic diagram of compression excitation.

[0068] Figure 5 A diagram illustrating the reweighting of the ACON activation function.

[0069] Figure 6 This is a schematic diagram of scale-enhanced fusion.

[0070] Figure 7 This is a schematic diagram of the detection results (regional comparison). Detailed Implementation

[0071] The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

[0072] like Figure 1-2As shown, a deep learning-based method for detecting rebar detail annotation text in engineering drawings includes the following steps:

[0073] S1. Preprocessing of engineering drawing images and annotation of text regions for steel reinforcement detail annotations;

[0074] The input engineering drawing images undergo format standardization preprocessing and image enhancement. This preprocessing includes format conversion (e.g., converting PDF or CAD drawings to JPG / PNG), enhancing edge information using the Sobel operator or Canny algorithm, and removing background noise using median filtering or bilateral filtering. This ensures that the images processed in subsequent recognition stages possess good contrast and boundary information, providing a foundation for text detection.

[0075] The annotation text region dataset for rebar detail drawings is labeled according to the prior layout rules of the rebar detail drawings in the engineering drawings. The semantic segmentation and annotation of the detail annotation region are performed manually. The annotation region is divided into three categories: numbering region, number of diameter region, and length region. Region labels are assigned to the logical relationship of each type of region annotation, which serves as a high-quality dataset for subsequent text detection network training and evaluation.

[0076] S2. Construct a multi-scale feature map extraction backbone network architecture in stages;

[0077] In the design of the text detection network, a lightweight convolutional neural network based on depthwise separable convolutions is constructed as the backbone network for text feature extraction. The framework design is as follows: Figure 3 As shown, this network is based on the MobileNet architecture and uses depthwise separable convolution to replace traditional convolution operations, which greatly reduces model parameters and computational load. While ensuring recognition capabilities, it achieves network lightweighting, making it easy to deploy to terminal devices and effectively reducing model computational complexity. At the same time, combined with staged multi-scale feature map extraction, the network has good detection capabilities for annotation text of different sizes and resolutions.

[0078] S3. Introduce compression excitation in the feature maps at each scale. Introduce compression excitation (SE) in the feature maps output by each network layer. Through global average pooling and inter-channel weight recalibration mechanism, automatically calculate the importance weight of each channel, thereby enhancing the channel response containing annotated text features and suppressing irrelevant background information, thereby improving the anti-interference ability.

[0079] The compression excitation consists of two parts, such as Figure 4As shown, in the compression part, global average pooling is used to condense the information of the entire feature map. The compressed feature vector can be regarded as a global information representation of the entire feature map, reflecting the importance of different channels in the feature map. In the activation part, two fully connected layers are used to learn the correlation between feature channels. First, a fully connected layer with a linear transformation and activation function is used to reduce the dimensionality of the global feature vector to a smaller dimension, and then a dimensionality-upgrading fully connected layer is used to map the dimensionality-down feature vector back to the original feature channel dimension.

[0080] S4. In nonlinear modeling, the ACON adaptive activation function is used, which endows neurons with the ability to learn whether to activate during nonlinear transformation, thereby improving the network's ability to distinguish complex backgrounds or blurred characters in drawings. ACON can self-regulate its activation state during training, effectively improving the network's expressive power and adaptability.

[0081] In nonlinear modeling, the ACON adaptive activation function is adopted, which breaks through the fixed nature of traditional activation functions and enables the model to flexibly determine whether neurons are activated, thereby adapting to the actual situation where characters are complex, densely arranged, and have a lot of line interference in the drawing.

[0082] By introducing a switching factor To learn about parameter switching between nonlinear and linear activation, we first unify the various forms of activation functions to obtain a general and simple adaptive activation function. Commonly used activation functions such as ReLUmax(x,0) can be represented or approximated as follows: The form in which It is a linear function. Therefore, this formula can be used to construct an adaptive activation function. For a maximum function By using the softmax approach for smooth approximation, we can obtain an approximate formula for the maximum value function:

[0083]

[0084] When n=2, the maximum value function It can be approximated as:

[0085]

[0086] Substituting the Sigmoid function into the above approximation formula, it can be expressed as:

[0087]

[0088] because and Both are linear functions, and we can let Where a and b are linear control coefficients that control the upper and lower bounds of the activation function and the activation result, we can obtain the general form of the ACON activation function:

[0089]

[0090] In the formula, Used to control whether a function is activated or not.

[0091] So, for input For the feature map, its output is An array of the same size as the input features, such as Figure 5 As shown. Feature map First, it passes through a mean filter, within each channel and The mean is applied across the dimensions, and then passed through one layer. and one layer The pointwise convolution is performed, and finally the Sigmoid activation function is used to obtain the input feature map. Same size This is used to control whether it is activated.

[0092] S5. Construct a multi-scale feature pyramid structure based on RSE-FPN and superimpose scale-enhanced fusion;

[0093] When fusing feature information, a multi-scale feature pyramid structure based on RSE-FPN is constructed, and scale-enhanced fusion is superimposed. By transposing convolution and channel alignment of feature maps at multiple scales, scale normalization and pixel-wise weighted fusion operations, the network's ability to distinguish text regions of different sizes is improved, enhancing the model's perception and expression of small characters and solving the key problem of missing single characters in engineering drawings.

[0094] This step constructs the neck of the neural network, integrating features extracted from the backbone network to generate a more representative and adaptive feature pyramid. The RSE-FPN network combines Res2Net (with multi-branch receptive field enhancement) and the Feature Pyramid Network (FPN). Res2Net is a convolutional neural network that enhances the network's receptive field by increasing the number of branches and channels within branches, and by processing features at different scales in parallel, thus improving the network's adaptability to scale changes.

[0095] Scale-enhanced fusion, implemented after the neck network, addresses the problem that as the size of the feature maps extracted by the backbone network decreases, the feature resolution also decreases, leading to the neglect or inaccurate detection of small areas of text. For example... Figure 6As shown, the features after concatenation at the network neck are divided into three branches: the first branch performs only one transposed convolution to upsample the features and then performs scale restoration through a 3×3 convolution to preserve the detection capability of the original text; the second branch performs two feature enhancement operations and then aligns the dimensions with the original features through pointwise convolution; the third branch expands the receptive field through two convolution operations and aligns the feature dimensions with the pointwise convolution through upsampling; finally, the feature maps of the three branches are superimposed to achieve scale enhancement fusion for text of different areas in the large sample annotation.

[0096] S6. By adopting a differentiable binarization strategy, an approximate binary map is output based on the obtained probability map and threshold map, and finally the accurate prediction of the bounding box is achieved.

[0097] Based on the fusion feature map obtained in step S5 The prediction is performed through two parallel branches. One branch predicts a probability map P for the text region, which reflects the likelihood of the text region in the image. The size of the probability map is the same as the original input image. Each pixel value in the probability map P is transformed using a differentiable activation function. To ensure the entire process is differentiable, the sigmoid function is used as the activation function.

[0098]

[0099] In the formula, This represents the probability value of the input. This represents the result after the Sigmoid function transformation, with values ​​ranging from (0,1), where k is the magnification factor. Another branch performs binarization threshold prediction, which is primarily used to guide the subsequent binarization process. The threshold calculation formula is as follows:

[0100]

[0101] In the formula, t is a predefined threshold, and (i,j) are the coordinate points in the graph;

[0102] The probability map after Sigmoid transformation is multiplied pixel by pixel with the threshold map to obtain an approximate binary map.

[0103] After obtaining the probability map, threshold map, and approximate binary map, cross-entropy loss is used as the loss function to measure the difference between the generated text probability map P and the ground truth label.

[0104] A probabilistic graph represents the probability that each pixel belongs to the text, helping to ensure that the network can accurately predict the location of text regions. The probabilistic graph cross-entropy loss function... The calculation formula is as follows:

[0105]

[0106] In the formula, 𝑁 represents the number of samples, and 𝐶 represents the number of categories. It is the One-Hot encoding of the real label. This represents the probability value at the corresponding position in the predicted probability map;

[0107] Threshold maps represent the pixel boundaries of text, helping to ensure that the network can learn appropriate boundary information. The threshold map cross-entropy loss function... The calculation formula is as follows:

[0108]

[0109] In the formula, This represents the threshold at the corresponding position in the predicted threshold map;

[0110] The approximate binary graph is the final output of the model, and its loss function is... as follows:

[0111]

[0112] In the formula, It is the pixel value at the corresponding position in the approximate binary image;

[0113] Finally, the total loss function is obtained by weighted summation of the probabilistic graphical loss, threshold graphical loss, and approximate binarization loss. The formula is shown below:

[0114]

[0115] In the formula, Let the probabilistic graphical cross-entropy loss function be used. The threshold graph cross-entropy loss function is used. This is the loss function for an approximate binary graph.

[0116] Figure 7 The diagram shows a comparison between the detection results of the present invention and traditional methods, demonstrating that the present invention has significant advantages in terms of circular number recognition, small character detection, and robustness in complex backgrounds.

[0117] This invention also provides a deep learning-based system for detecting text in reinforcement detail annotations on engineering drawings, comprising the following modules:

[0118] - Image preprocessing and dataset annotation module, used for preprocessing engineering drawing images and annotating the text regions of steel bar detail annotations;

[0119] - Backbone network architecture building module, used to build a lightweight convolutional neural network based on depthwise separable convolutions as the backbone network for text feature extraction;

[0120] - The compression stimulus module is used to automatically calculate the importance weight of each channel through global average pooling and inter-channel weight recalibration mechanism, enhance the channel response containing annotated text features, and suppress irrelevant background information;

[0121] - Nonlinear modeling module, used to introduce ACON adaptive activation function, to give neurons the ability to learn whether to activate in the nonlinear transformation stage, to improve the network’s ability to distinguish complex backgrounds or blurry characters in drawings. ACON can self-adjust activation state during training, effectively improving the network’s expressive ability and adaptability.

[0122] - The pyramid structure construction and scale enhancement fusion module is used to improve the network's ability to distinguish text regions of different sizes and enhance the model's perception and expression of small characters by performing channel alignment, scale normalization and pixel-wise weighted fusion operations on feature maps at multiple scales.

[0123] - The text detection and prediction module is used to output an approximate binary map based on the obtained probability map and threshold map, and finally achieve accurate prediction of the bounding box.

[0124] The present invention also provides an electronic device, which includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other via the communication bus.

[0125] Memory, used to store computer programs;

[0126] The processor executes the computer program stored in the memory to implement the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings, as described above.

[0127] The present invention also provides a non-volatile storage medium storing an executable program, which, when executed by a processor, implements the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings as described above.

[0128] The above specific embodiments are used to explain and illustrate the present invention, and are only preferred embodiments of the present invention, not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for detecting detailed reinforcement annotation text in engineering drawings based on deep learning, characterized in that: Includes the following steps: S1. Preprocessing of engineering drawing images and annotation of text regions for steel reinforcement detail annotations; S2. Construct a multi-scale feature map extraction backbone network architecture in stages; S3. Introduce compression excitation in the feature maps at each scale; S4. The ACON adaptive activation function is used for nonlinear modeling. S5. Construct a multi-scale feature pyramid structure based on RSE-FPN and superimpose scale-enhanced fusion; S6. Output an approximate binary map to achieve accurate prediction of the bounding box.

2. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S1, the preprocessing of the engineering drawing image includes image format conversion, using the Sobel operator or Canny algorithm to enhance edge information, and using median filtering or bilateral filtering to remove noise interference in the image background.

3. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S1, the annotation of the rebar detail annotation text region dataset includes manual semantic segmentation annotation of the detail annotation region, dividing the annotation region into three categories: number region, number diameter region, and length region, and assigning region labels to the logical relationship of each type of region annotation.

4. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S3, compression is achieved by global average pooling to condense the information of the entire feature map, and excitation is achieved by two fully connected layers to learn the correlation between feature channels.

5. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S4, a switching factor is introduced. To learn the parameter switching between nonlinear and linear activation, for a maximum function By using the softmax approach for smooth approximation, we can obtain an approximate formula for the maximum value function: When n=2, the maximum value function It can be approximated as: Substituting the Sigmoid function into the above approximation formula, it can be expressed as: because and Both are linear functions, and we can let Where a and b are linear control coefficients that control the upper and lower bounds of the activation function and the activation result, we can obtain the general form of the ACON activation function: In the formula, Used to control whether a function is activated or not.

6. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S5, the RSE-FPN network is a combination of Res2Net with multi-branch receptive field enhancement and FPN with feature pyramid. The features after being concatenated at the network neck are divided into three branches: the first branch performs only one transposed convolution to upsample the features and performs scale restoration through 3×3 convolution to preserve the detection capability of the original text; the second branch performs two feature enhancement operations and aligns the dimensions with the original features through pointwise convolution; the third branch expands the receptive field through two convolution operations and aligns the feature dimensions with the pointwise convolution through upsampling; finally, the feature maps of the three branches are superimposed to achieve scale enhancement fusion for text of different areas in large sample annotations.

7. The method for detecting reinforcement detail annotation text in engineering drawings based on deep learning according to claim 1, characterized in that: In step S6, each pixel value in the probability map P is transformed using a differentiable activation function. To ensure the entire process is differentiable, the Sigmoid function is used as the activation function. In the formula, This represents the probability value of the input. This represents the result after the Sigmoid function transformation, with values ​​ranging from (0,1), where k is the magnification factor; the other branch performs binarized threshold map prediction, and the threshold map is calculated using the following formula: In the formula, t is a predefined threshold, and (i,j) are the coordinate points in the graph; The probability map after Sigmoid transformation is multiplied pixel by pixel with the threshold map to obtain an approximate binary map. After obtaining the probability map, threshold map, and approximate binary map, cross-entropy loss is used as the loss function. Probabilistic graphical cross-entropy loss function The calculation formula is as follows: In the formula, 𝑁 represents the number of samples, and 𝐶 represents the number of categories. It is the One-Hot encoding of the real label. This represents the probability value at the corresponding position in the predicted probability map; Threshold graph cross-entropy loss function The calculation formula is as follows: In the formula, This represents the threshold at the corresponding position in the predicted threshold map; Approximate binary graph loss function The calculation formula is as follows: In the formula, It is the pixel value at the corresponding position in the approximate binary image; Finally, the total loss function is obtained by weighted summation of the probabilistic graphical loss, threshold graphical loss, and approximate binarization loss. The formula is shown below: In the formula, Let the probabilistic graphical cross-entropy loss function be used. The threshold graph cross-entropy loss function is used. This is the loss function for an approximate binary graph.

8. A deep learning-based system for detecting text in engineering drawing reinforcement detail annotations, characterized in that: Includes the following modules: - Image preprocessing and dataset annotation module, used for preprocessing engineering drawing images and annotating the text regions of steel bar detail annotations; - Backbone network architecture building module, used to build a lightweight convolutional neural network based on depthwise separable convolutions as the backbone network for text feature extraction; - The compression stimulus module is used to automatically calculate the importance weight of each channel through global average pooling and inter-channel weight recalibration mechanism, enhance the channel response containing annotated text features, and suppress irrelevant background information; - Nonlinear modeling module, used to introduce ACON adaptive activation function, to give neurons the ability to learn whether to activate in the nonlinear transformation stage, to improve the network’s ability to distinguish complex backgrounds or blurry characters in drawings. ACON can self-adjust activation state during training, effectively improving the network’s expressive ability and adaptability. - The pyramid structure construction and scale enhancement fusion module is used to improve the network's ability to distinguish text regions of different sizes and enhance the model's perception and expression of small characters by performing channel alignment, scale normalization and pixel-wise weighted fusion operations on feature maps at multiple scales. - The text detection and prediction module is used to output an approximate binary map based on the obtained probability map and threshold map, and finally achieve accurate prediction of the bounding box.

9. An electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus, characterized in that: The memory is used to store computer programs; A processor, the processor being configured to execute a computer program stored in a memory to implement the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings as described in any one of claims 1-7.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium stores an executable program, which, when executed by a processor, implements the steps of the deep learning-based method for detecting reinforcement detail annotation text in engineering drawings as described in any one of claims 1-7.

Images