Building surface disease identification method based on space-frequency-edge multi-domain fusion
By employing a spatial-frequency-edge multi-domain fusion method, and utilizing convolutional neural networks and multi-domain fusion modules, the problem of insufficient accuracy in disease identification in complex scenarios in existing technologies has been solved, achieving efficient and stable identification of building surface diseases.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-17
AI Technical Summary
Existing methods for identifying surface defects in buildings lack accuracy in complex scenarios. Single spatial domain features are easily affected by noise, and frequency domain features are difficult to accurately locate defect areas. Furthermore, the lack of an effective multi-domain fusion mechanism results in insufficient detection accuracy and robustness.
A spatial-frequency-edge multi-domain fusion method is adopted, which extracts spatial, frequency and edge features of building surface images through convolutional neural networks, and uses a multi-domain fusion module for feature fusion, including block flattening, sequence construction and multi-head attention fusion, adaptively weighting features from different domains to enhance the accuracy and stability of disease identification.
It significantly improves the accuracy and robustness of disease identification in complex scenarios such as changes in lighting, surface contamination, and rough textures, enabling more precise detection of minute cracks and diseased areas, and improving the model's performance in complex engineering scenarios.
Smart Images
Figure CN121884128A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to a method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion. Specifically, this invention utilizes convolutional neural networks, frequency domain feature extraction, edge structure feature enhancement, and multi-domain feature fusion mechanisms to automatically identify and detect defects such as cracks and spalling on building surfaces such as concrete and walls, falling within the technical scope of building structural health monitoring and intelligent detection. Background Technology
[0002] Buildings, bridges, tunnels, and other structures often develop surface defects such as cracks, spalling, weathering, and voids during long-term use. Timely and accurate identification of these defects is crucial for ensuring building safety and extending structural lifespan. Traditional manual inspection methods suffer from low efficiency, high subjectivity, and high safety risks, making them unsuitable for the large-scale and high-frequency inspection needs of modern engineering projects.
[0003] To address these issues, image-based automated disease detection methods have gradually become a research hotspot. Early studies mostly employed traditional image processing algorithms, such as Canny edge detection, threshold segmentation, grayscale variation analysis, and morphological processing, identifying crack areas through manually set rules. However, due to the complex textures of building surfaces, variable lighting, and severe interference from stains, these methods are susceptible to noise, exhibiting weak robustness and failing to meet practical requirements in terms of detection accuracy.
[0004] With the development of deep learning, convolutional neural networks and Transformer-type models have been introduced into the detection of surface defects in buildings. Related research shows that spatial domain features can effectively extract the appearance and texture information of cracks, frequency domain features can enhance the representation of structural details and high-frequency textures, while edge domain information plays an important role in the geometry of cracks and the localization of slender structures. At the same time, the comprehensive analysis of multi-domain information helps to improve the accuracy and stability of defect identification.
[0005] In the process of realizing this invention, the following problems were found in the prior art:
[0006] Single spatial domain features are difficult to adapt to complex scenarios. Under conditions such as rough surfaces, shadow interference, and stain coverage, spatial domain features are easily affected by noise, leading to false detections or missed detections; frequency domain features cannot directly reflect the morphological structure of defects such as cracks, making it difficult to accurately locate defect areas; research on the effective integration of edge information with other domains is relatively insufficient, resulting in the value of edge information not being fully utilized.
[0007] There is a lack of effective multi-domain fusion mechanisms. Most existing methods are still at the stage of single-domain or dual-domain feature extraction, lacking deep fusion modules that can adaptively fuse spatial, frequency, and edge domains, resulting in the failure to fully leverage the complementarity between multi-domain features.
[0008] Therefore, a method is still needed that can comprehensively integrate spatial domain texture information, frequency domain high-frequency details, and edge geometric features to improve the accuracy, robustness, and adaptability of building surface defect identification. Summary of the Invention
[0009] To address the shortcomings of existing building surface defect identification methods, such as insufficient accuracy in complex scenarios, inadequate extraction of crack geometry, and weak multi-domain feature fusion capabilities, this invention proposes a building surface defect identification method based on spatial-frequency-edge multi-domain fusion. This method fully leverages the complementarity between features from different domains, ultimately improving the detection effectiveness and robustness of building surface defects.
[0010] To achieve the above objectives, this invention provides a method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion, comprising the following steps:
[0011] Step S1: Acquire images of the building surface and perform preprocessing to obtain images to be classified;
[0012] Step S2: Extract spatial domain features from the image to be classified;
[0013] Step S3: Extract the frequency domain features of the image to be classified;
[0014] Step S4: Extract edge region features from the image to be classified;
[0015] Step S5: Input spatial domain features, frequency domain features, and edge domain features into the multi-domain fusion module for multi-domain feature fusion;
[0016] Step S6: Input the fused features into the classifier and output the disease identification results.
[0017] In a further optimization of this technical solution, step S1 includes:
[0018] (1) Obtain images of building surfaces. Image sources include: images collected by engineering inspection drones, images collected by inspection robots, images captured by handheld devices, and existing monitoring images from data centers;
[0019] (2) Preprocessing the image, including applying one or more of the following processes to the image: size normalization, rotation, brightness adjustment, contrast adjustment, random cropping, etc., to obtain the image to be classified.
[0020] In a further optimization of this technical solution, in step S2, the image to be classified is input into a spatial domain feature extraction network to obtain spatial domain features.
[0021] In a further optimization of this technical solution, step S3 involves performing a frequency domain transformation on the image to be classified to obtain a frequency domain image, and then inputting the frequency domain image into a frequency domain feature extraction network to obtain frequency domain features.
[0022] This technical solution is further optimized by using lightweight convolutional neural networks, EfficientNet, or other backbone networks with efficient feature representation capabilities for the spatial domain feature extraction network and the frequency domain feature extraction network.
[0023] In a further optimization of this technical solution, step S4 involves edge detection of the image to be classified to obtain an edge image. This edge detection includes the Canny algorithm and... The algorithm fuses data through stacking:
[0024]
[0025] in For edge images, Stack represents the stacking method, Canny represents the edges detected by the Canny algorithm, and Sobel represents the edges detected by the Canny algorithm. x and Sobel y They are respectively The algorithm detects vertical and horizontal edges;
[0026] The edge image is then input into the edge domain feature extraction network to obtain edge domain features.
[0027] This technical solution is further optimized in that the edge domain feature extraction network is a lightweight convolutional neural network consisting of four convolutional layers, including convolutional layers, batch normalization layers, and ReLU activation layers, used to extract the geometric morphological features of cracks:
[0028]
[0029] Where F e Features of the marginal domain This is a feature extraction network for edge regions.
[0030] This technical solution is further optimized by step S5, which involves the spatial domain feature F. s Frequency domain characteristics F f and marginal domain features F e The input multi-domain fusion module adopts the following structure:
[0031] (1) Cut into pieces and flatten.
[0032] Image features in the spatial domain, frequency domain, and edge domain are divided into patches of uniform size:
[0033]
[0034] Where P s P f and P e These are feature patches for the spatial domain, frequency domain, and edge domain, respectively. It is a two-dimensional convolution.
[0035] The feature blocks obtained after flattening:
[0036]
[0037] Where T s T f and T e These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with Flatten being the flattening operation.
[0038] (2) Sequence construction
[0039] Embed learnable positional codes into one-dimensional feature blocks in the spatial, frequency, and edge domains:
[0040]
[0041] Where T s '、T f 'and T e 'These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with positional encoding.' , and These are learnable positional codes in the spatial domain, frequency domain, and edge domain, respectively.
[0042] (3) Multi-head attention fusion
[0043] The obtained one-dimensional feature block with positional encoding is input into the Transformer encoder, and global interaction and adaptive fusion are achieved using multi-head self-attention:
[0044]
[0045] Where F fusion As a feature of fusion, MHA is a multi-head attention mechanism.
[0046] This technical solution is further optimized by step S6, which involves fusing the features F. fusion Input the classifier and output the disease identification results.
[0047]
[0048] Where Y represents the disease identification result, and Classify is a multi-layer fully connected classifier.
[0049] Further optimization of this technical solution includes adding a geometric structure guidance module before multi-domain feature fusion in step S5, using edge maps to generate geometric attention masks to guide spatial and frequency domain features to focus on possible crack regions:
[0050]
[0051] Where M is the geometric attention map and G is the geometric structure guidance function, whose function is to process the features F extracted from the edge region. e Convert to a geometric attention map;
[0052] Use M as the geometric weight, along with F s and F f Perform element-wise weighted summation:
[0053]
[0054] in and These are the weighted spatial domain and frequency domain features, respectively.
[0055] The obtained weighted spatial domain and frequency domain features can replace the original F s and F f Used for multi-domain feature fusion.
[0056] The above technical solution has the following advantages, unlike existing technologies:
[0057] Integrating features from three domains to improve robustness: Features are extracted simultaneously in the spatial, frequency, and edge domains, making full use of complementary information between different domains to significantly improve recognition capabilities in complex scenarios such as changes in lighting, surface contamination, and rough textures.
[0058] Highlighting the geometric structure of cracks: Edge domain networks can more accurately describe the linear, elongated, and topologically continuous characteristics of cracks, making the model perform better in detecting fine cracks.
[0059] Frequency domain enhancement of detail representation: Frequency domain texture is extracted by discrete cosine transform, which effectively enhances the contribution of high-frequency details in disease identification and improves the ability to distinguish between small cracks and background noise.
[0060] Multi-domain deep fusion mechanism: Through adaptive weighting or attention-based multi-domain fusion modules, more effective feature interaction is achieved, making the overall model more stable in complex engineering scenarios.
[0061] In summary, this invention can accurately identify the types of defects on building surfaces under various complex conditions, and has high engineering application value and practical significance. Attached Figure Description
[0062] Figure 1 The overall flow of the invention method described in the specific implementation embodiment;
[0063] Figure 2 The specific implementation method describes step S1, image acquisition and preprocessing.
[0064] Figure 3 The specific implementation method describes step S2, spatial domain feature extraction.
[0065] Figure 4 The specific implementation method describes step S3, frequency domain feature extraction.
[0066] Figure 5 Step S4, edge region feature extraction, is described in the specific implementation method.
[0067] Figure 6 Step S5, multi-domain feature fusion, is described in the specific implementation method.
[0068] Figure 7 The specific implementation method includes step S6, disease identification.
[0069] Explanation of reference numerals in the attached figures:
[0070] In the attached diagram, the symbols starting with "S" represent steps, the symbols starting with "M" represent functional modules, and the symbols starting with "D" represent data objects.
[0071] S101. Obtain an image of the building surface.
[0072] M101, Image Preprocessing
[0073] D101, Image to be classified.
[0074] M201, Spatial Domain Feature Extraction Module
[0075] D201, Spatial Domain Characteristics.
[0076] M301, Frequency Domain Transformation Module
[0077] D301, frequency domain image,
[0078] M302, Frequency Domain Feature Extraction Module
[0079] D302, Frequency Domain Characteristics.
[0080] M401, edge detection module,
[0081] D401, edge image,
[0082] M402, Edge Region Feature Extraction Module
[0083] D402, marginal domain features.
[0084] M501, Segmentation Embedding Module
[0085] M502, Multi-head Attention Fusion Module
[0086] D501, Fusion Features.
[0087] M601, classifier,
[0088] D601, Disease identification results.
[0089] S101, Image Acquisition and Preprocessing;
[0090] S102, Spatial Domain Feature Extraction;
[0091] S103, Frequency Domain Feature Extraction;
[0092] S104, Edge region feature extraction;
[0093] S105, Multi-domain feature fusion;
[0094] S106. Disease identification. Detailed Implementation
[0095] To illustrate the technical content, structural features, objectives, and effects of the technical solution in detail, the following description is provided in conjunction with specific embodiments and accompanying drawings. It should be noted that those skilled in the art can modify or substitute the following embodiments without altering the principles of the invention, and such modifications and substitutions all fall within the scope of protection of this invention.
[0096] Example 1: A method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion
[0097] See Figure 1 The diagram shows the overall flow of the invention method described in a specific embodiment. This embodiment provides a multi-domain fusion identification method for detecting defects such as cracks and peeling on building surfaces. The method includes the following steps:
[0098] Step S1: Image Acquisition and Preprocessing
[0099] like Figure 2 The diagram shown illustrates the image acquisition and preprocessing process.
[0100] (1) Obtain images of building surfaces, including but not limited to:
[0101] Images captured by engineering inspection drones
[0102] Inspection robot collects images
[0103] Images taken with handheld devices
[0104] Data center already has monitoring images
[0105] (2) Preprocessing the image, including but not limited to:
[0106] Normalize the image size to 224×224 or other set sizes;
[0107] Optional data augmentations, such as rotation, brightness adjustment, contrast adjustment, and random cropping, can be used to improve model robustness.
[0108] The images to be classified are obtained through the above steps.
[0109] Step S2: Spatial Domain Feature Extraction
[0110] Spatial domain feature extraction: After preprocessing the input building surface image, a convolutional neural network or a lightweight visual backbone network is used to extract spatial domain texture features to describe surface texture, brightness changes and appearance information of defects.
[0111] like Figure 3 As shown, the image to be classified is input into the spatial domain feature extraction network.
[0112] In an optional embodiment, the spatial domain feature extraction network is a lightweight convolutional neural network, EfficientNet, or other backbone network with efficient feature representation capabilities.
[0113] In this embodiment, a pre-trained EfficientNetV2-S network is used to extract spatial domain features:
[0114]
[0115] Where F s For spatial domain characteristics, To pre-train the EfficientNetV2-S network, The input image.
[0116] This feature is mainly used to describe the brightness variations, texture features, and appearance features of diseases in an image.
[0117] Step S3: Frequency Domain Feature Extraction
[0118] Frequency domain feature extraction: The input image is subjected to frequency domain transformations such as discrete cosine transform to extract its high-frequency and mid-frequency components, so as to enhance the ability to express local details such as crack edges and high-frequency texture changes, and further learn frequency domain features through deep networks.
[0119] like Figure 4As shown, the image to be classified is transformed in the frequency domain to obtain the frequency domain image.
[0120] In an optional embodiment, the frequency domain transformation module is a discrete cosine transform:
[0121]
[0122] in The image is after discrete cosine transform. This is the discrete cosine transform method.
[0123] The frequency domain image is input into the frequency domain feature extraction network.
[0124] In an optional embodiment, the frequency domain feature extraction network is a lightweight convolutional neural network, EfficientNet, or other backbone network with efficient feature representation capabilities.
[0125]
[0126] Where F f It is a feature of the spatial domain.
[0127] This feature enhances high-frequency details and local structural variations, helping to distinguish crack textures from background noise.
[0128] Step S4: Edge Region Feature Extraction
[0129] Edge domain feature extraction: Edge detection algorithms are used to generate edge images to highlight the geometric features of defects such as cracks, and edge domain structural features are extracted through a lightweight convolutional network.
[0130] like Figure 5 As shown, edge detection is performed on the image to be classified to obtain the edge image.
[0131] In an optional embodiment, the edge detection module comprises the Canny algorithm and The algorithm uses a stacking method for fusion:
[0132]
[0133] in For edge images, Stack represents the stacking method, Canny represents the edges detected by the Canny algorithm, and Sobel represents the edges detected by the Canny algorithm. x and Sobel y They are respectively The algorithm detects vertical and horizontal edges.
[0134] The edge image is then input into the edge domain feature extraction network.
[0135] In an optional embodiment, the edge region feature extraction network is a lightweight convolutional neural network consisting of four convolutional layers, including convolutional layers, batch normalization layers, and ReLU activation layers, used to extract crack geometric features:
[0136]
[0137] Where F e Features of the marginal domain This is a feature extraction network for edge regions.
[0138] This feature can highlight the linear structure, boundary continuity, and topological morphology of cracks.
[0139] Step S5: Multi-domain feature fusion
[0140] Multi-domain fusion processing: Construct a deep fusion module of spatial domain, frequency domain and edge domain, and use position encoding and attention mechanism to adaptively weight and fuse the three types of features to obtain a more comprehensive and discriminative disease expression.
[0141] Step S5 is the core innovative module of this invention. Its purpose is to achieve deep collaboration and adaptive weighting among spatial domain features, frequency domain features, and edge domain features at the feature representation level, so as to fully leverage the complementarity of the three domain features and improve the ability to identify slender cracks, minor defects, and complex backgrounds. This step can introduce global interaction while preserving local details, so that high-frequency information in the frequency domain, geometric information of the edges, and spatial texture are fully complementary at the representation level, enhancing crack / defect signals and suppressing background noise. Figure 6 As shown, this step will transform the spatial domain features F s Frequency domain characteristics F f and marginal domain features F e The input is processed by a multi-domain fusion module, and the fused feature F is obtained through steps such as attention fusion. fusion .
[0142] In an optional embodiment, the multi-domain fusion module adopts the following structure:
[0143] (1) Cut into pieces and flatten.
[0144] Image features in the spatial domain, frequency domain, and edge domain are divided into patches of uniform size:
[0145]
[0146] Where P s P f and P e These are feature patches for the spatial domain, frequency domain, and edge domain, respectively. It is a two-dimensional convolution.
[0147] The feature blocks obtained after flattening:
[0148]
[0149] Where T s T f and T e These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with Flatten being the flattening operation.
[0150] (2) Sequence construction
[0151] Embed learnable positional codes into one-dimensional feature blocks in the spatial, frequency, and edge domains:
[0152]
[0153] Where T s '、T f 'and T e 'These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with positional encoding.' , and These are learnable positional codes in the spatial domain, frequency domain, and edge domain, respectively.
[0154] (3) Multi-head attention fusion
[0155] The obtained one-dimensional feature block with positional encoding is input into the Transformer encoder, and global interaction and adaptive fusion are achieved using multi-head self-attention:
[0156]
[0157] Where F fusion As a feature of fusion, MHA is a multi-head attention mechanism.
[0158] This step enables the model to adaptively highlight the domain features that best represent the disease in different scenarios.
[0159] Step S6: Disease Identification
[0160] Disease identification and output: Through fully connected layers and classifiers, the system identifies the types or areas of diseases such as cracks and peeling on building surfaces and outputs the detection results.
[0161] like Figure 7 As shown, the fused features F fusion Input the classifier and output the disease identification results.
[0162]
[0163] Where Y represents the disease identification result, and Classify is a multi-layer fully connected classifier.
[0164] Example 2: Optional geometry guidance module
[0165] In another implementation, a geometric structure guidance module can be added before the S5 multi-domain feature fusion to generate a geometric attention mask using the edge map, guiding spatial and frequency domain features to focus on possible crack regions.
[0166]
[0167] Where M is the geometric attention map and G is the geometric structure guidance function, whose function is to process the features F extracted from the edge region. e Convert to a geometric attention graph.
[0168] Use M as the geometric weight, along with F s and F f Perform element-wise weighted summation:
[0169]
[0170] in and These are the weighted spatial domain and frequency domain features, respectively.
[0171] The obtained weighted spatial domain and frequency domain features can replace the original F s and F f Used for multi-domain feature fusion.
[0172] This module can further improve the accuracy of crack detection.
[0173] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element. Additionally, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number.
[0174] Although the above embodiments have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion, characterized in that, Includes the following steps: Step S1: Acquire images of the building surface and perform preprocessing to obtain images to be classified; Step S2: Extract spatial domain features from the image to be classified; Step S3: Extract the frequency domain features of the image to be classified; Step S4: Extract edge region features from the image to be classified; Step S5: Input spatial domain features, frequency domain features, and edge domain features into the multi-domain fusion module for multi-domain feature fusion; Step S6: Input the fused features into the classifier and output the disease identification results.
2. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, Step S1 includes: (1) Obtain images of building surfaces. Image sources include: images collected by engineering inspection drones, images collected by inspection robots, images captured by handheld devices, and existing monitoring images from data centers; (2) Preprocessing the image, including applying one or more of the following processes to the image: size normalization, rotation, brightness adjustment, contrast adjustment, random cropping, etc., to obtain the image to be classified.
3. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, In step S2, the image to be classified is input into the spatial domain feature extraction network to obtain spatial domain features.
4. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, Step S3 involves performing a frequency domain transformation on the image to be classified to obtain a frequency domain image. The frequency domain image is then input into a frequency domain feature extraction network to obtain frequency domain features.
5. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 3 or 4, characterized in that, The spatial domain feature extraction network and the frequency domain feature extraction network are lightweight convolutional neural networks, EfficientNet, or other backbone networks with efficient feature representation capabilities.
6. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, In step S4, edge detection is performed on the image to be classified to obtain an edge image. The edge detection includes the Canny algorithm and... The algorithm fuses data through stacking: in For edge images, Stack represents the stacking method, Canny represents the edges detected by the Canny algorithm, and Sobel represents the edges detected by the Canny algorithm. x and Sobel y They are respectively The algorithm detects vertical and horizontal edges; then the edge images are input into the edge domain feature extraction network to obtain edge domain features.
7. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 6, characterized in that, The edge region feature extraction network is a lightweight convolutional neural network consisting of four convolutional layers, including convolutional layers, batch normalization layers, and ReLU activation layers, used to extract crack geometric features. Where F e Features of the marginal domain This is a feature extraction network for edge regions.
8. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, Step S5 will define the spatial domain feature F s Frequency domain characteristics F f and marginal domain features F e The input multi-domain fusion module adopts the following structure: (1) Cut into pieces and flatten. Image features in the spatial domain, frequency domain, and edge domain are divided into patches of uniform size: Where P s P f and P e These are feature patches for the spatial domain, frequency domain, and edge domain, respectively. It is a two-dimensional convolution. The feature blocks obtained after flattening: Where T s T f and T e These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with Flatten being the flattening operation. (2) Sequence construction Embed learnable positional codes into one-dimensional feature blocks in the spatial, frequency, and edge domains: Where T s '、T f 'and T e 'These are one-dimensional feature blocks in the spatial domain, frequency domain, and edge domain, respectively, with positional encoding.' , and These are learnable positional codes in the spatial domain, frequency domain, and edge domain, respectively. (3) Multi-head attention fusion The obtained one-dimensional feature block with positional encoding is input into the Transformer encoder, and global interaction and adaptive fusion are achieved using multi-head self-attention: Where F fusion As a feature of fusion, MHA is a multi-head attention mechanism.
9. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, Step S6 will fuse the feature F fusion Input the classifier and output the disease identification results. Where Y represents the disease identification result, and Classify is a multi-layer fully connected classifier.
10. The method for identifying building surface defects based on spatial-frequency-edge multi-domain fusion as described in claim 1, characterized in that, It also includes a geometric structure guidance module added before the multi-domain feature fusion in step S5, which uses the edge map to generate a geometric attention mask to guide the spatial and frequency domain features to focus on possible crack regions: Where M is the geometric attention map and G is the geometric structure guidance function, whose function is to process the features F extracted from the edge region. e Convert to a geometric attention map; Use M as the geometric weight, along with F s and F f Perform element-wise weighted summation: in and These are the weighted spatial domain and frequency domain features, respectively. The obtained weighted spatial domain and frequency domain features can replace the original F s and F f Used for multi-domain feature fusion.