Concrete disease detection method based on omnidirectional cross scanning mechanism
The pier defect segmentation network using an omnidirectional cross-scanning mechanism solves the problems of high computational cost and insufficient robustness of the Transformer model in high-resolution image detection, achieving efficient and accurate concrete defect detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-03-27
AI Technical Summary
Existing Transformer-based concrete defect detection models suffer from high computational cost, slow speed, and high GPU memory consumption during high-resolution image inference. They also suffer from limited receptive field, performance degradation, and insufficient interpretability.
The bridge pier defect segmentation network, which adopts an omnidirectional cross-scanning mechanism, includes an encoder and a decoder. It uses an omnidirectional state space module (OSS Block) for feature enhancement and extraction, and combines a fully connected layer for segmentation to achieve efficient defect detection.
It achieves accurate detection of high-resolution concrete images, improves the accuracy of edge detail detection, reduces the number of floating-point operations in the model when predicting defects, and enhances robustness and stability.
Smart Images
Figure CN121746908A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automated detection technology for apparent defects in infrastructure concrete, specifically a method for detecting concrete defects based on an omnidirectional cross-scanning mechanism. Background Technology
[0002] In recent years, deep learning-based image processing technology has been widely applied in the field of concrete defect detection due to its high performance, flexibility, and robustness. Detection methods are mainly divided into two categories: models based on convolutional neural networks (CNNs) and models based on Transformers. CNN-based models perform well in bridge pier defect detection tasks because they can effectively capture local spatial features through hierarchical structures. However, due to their limited ability to capture the global receptive field, CNNs struggle to extract comprehensive global spatial features and dependencies from bridge pier defect images.
[0003] Conversely, Transformer overcomes the limitations of CNNs in capturing global spatial relationships by using a self-attention mechanism to model the global context of the entire image. Du et al. proposed an efficient crack segmentation model that combines a self-attention mechanism with a weighted bidirectional feature pyramid, and combined it with image processing techniques to perform fine-grained crack segmentation, achieving a crack width quantization error within 0.1 mm.
[0004] While the global modeling capabilities of the Transformer model make it excellent for concrete defect detection tasks, the computational complexity of its self-attention mechanism not only dramatically increases the computational cost and slows down the model when inferring high-resolution defect images, but also consumes a large amount of GPU memory. Although some works have effectively improved the computational complexity of the second-order mechanism, they generally suffer from drawbacks such as limited receptive field, performance degradation, and insufficient interpretability. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a concrete defect detection method based on an omnidirectional cross-scanning mechanism, which achieves accurate and efficient detection of concrete surface defects, while possessing linear computational complexity. When inferring high-resolution images, it is more computationally efficient than the Transformer model and exhibits superior robustness.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for detecting concrete defects based on an omnidirectional cross-scanning mechanism, specifically including the following steps: Step 1: Obtain the image dataset of apparent defects in the concrete of the bridge piers and preprocess it; Step 2: Label the preprocessed image dataset to obtain the labeled image dataset. Randomly divide the labeled image dataset into training set and test set according to the proportion. The labeled image dataset includes segmentation information of multiple bridge pier surface defects, including background images, crack images, exposed rebar images and peeling images. Step 3: Construct a bridge pier defect segmentation network based on an omnidirectional cross-scanning mechanism; Step 4: Using the training set obtained in Step 2 as input and the segmentation information of the apparent defects of the bridge piers as output, train the bridge pier defect segmentation network based on the omnidirectional cross-scanning mechanism obtained in Step 3 to obtain the bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism. Step 5: Input the test set obtained in Step 2 into the pier defect segmentation model based on the omnidirectional cross-scanning mechanism obtained in Step 4 for segmentation to obtain the segmented test set, which includes the apparent defect results of multiple pier concrete.
[0007] Preferably, the bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism includes an encoder and a decoder connected in sequence. The encoder is used to extract multi-scale feature information of the bridge pier apparent defect image, and performs feature enhancement and extraction through multiple omnidirectional state space modules (OSS Blocks) to output multiple feature maps. The decoder is used to perform feature fusion and enhancement on the multiple feature maps output by the encoder through a feature fusion network, and simultaneously uses a fully connected layer to segment the processed feature maps and output the segmentation result.
[0008] Preferably, the encoder includes four stages connected in sequence: the first stage includes a connected patch embedding module and two OSS Block modules; the second stage includes a connected patch merging module and two OSS Block modules; the third stage includes a connected patch merging module and eight OSS Block modules; and the fourth stage includes a connected patch merging module and two OSS Block modules.
[0009] Preferably, the OSS Block module includes a first part and a second part connected in series; the first part includes a first-layer normalization module LN and an omnidirectional selection scanning module OSSM connected in sequence, and the output of the omnidirectional selection scanning module OSSM is also residually connected to the input of the first-layer normalization module LN; the second part includes a second-layer normalization module LN and a feedforward neural network FFN module connected in sequence, and the output of the feedforward neural network FFN module is also residually connected to the input of the second-layer normalization module LN.
[0010] Preferably, the OSSM module includes a multilayer perceptron module (MLP), a depthwise separable convolutional model (DWConv), a SiLU activation function, an omnidirectional selection scanning mechanism (OSS) module, a layer normalization module (LN), and the multilayer perceptron module (MLP) connected in series.
[0011] Preferably, the omnidirectional selection scanning mechanism OSS module includes a serially connected scan extension operation module, SSM module, and scan merging module; the scan extension operation module is used to perform cross-scanning of the input image in eight directions (horizontal, vertical, diagonal, and anti-diagonal) to generate eight different image block sequences; the scan merging module sums the eight sequences output by the SSM module, merges the global modeling information of multiple directions, and restores the output image to the same shape as the input.
[0012] Preferably, the decoder sequentially passes all feature maps output by the encoder through the MLP unified feature channel, and then obtains the fused feature map through Concat feature fusion. The fused feature map is then adjusted by the MLP to adjust the number of channels to the number of disease categories. Finally, the final segmented image is generated by upsampling to obtain the apparent disease results of the bridge pier concrete.
[0013] Preferably, the preprocessing includes data augmentation using geometric distortion and photometric distortion methods.
[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the concrete defect detection method based on an omnidirectional cross-scanning mechanism disclosed in this application.
[0015] A computer program product includes a computer program / instructions that, when executed by a processor, implement the concrete defect detection method based on an omnidirectional cross-scanning mechanism disclosed in this application.
[0016] Compared with the prior art, the beneficial technical effects of this invention are: (1) The present invention provides a concrete defect detection method based on an omnidirectional cross-scanning mechanism, which can achieve accurate detection of apparent defects in high-resolution concrete images under complex backgrounds, improve the detection accuracy of three defects on concrete surface: cracks, exposed reinforcement and spalling, and is particularly outstanding in the detection of some edge details.
[0017] (2) The concrete disease detection method based on the omnidirectional cross-scanning mechanism of the present invention can realize the implementation reasoning of high-resolution disease images and reduce the number of floating-point operations required by the model when predicting diseases.
[0018] (3) The concrete defect detection method based on the omnidirectional cross-scanning mechanism of the present invention can extract and retain key features more stably when the input defect image is subjected to various degrees of interference, and has better robustness. Attached Figure Description
[0019] Figure 1 This is a diagram showing the overall structure of a concrete defect detection algorithm based on an omnidirectional cross-scanning mechanism.
[0020] Figure 2 The diagram shows the omnidirectional state space module and its internal structure; where 2(a) is OSS Bloc, 2(b) is OSSModule, and 2(c) is Omnidirectional Selective Scan.
[0021] Figure 3 The diagram illustrates the scanning methods; where 3(a) represents horizontal forward scanning, 3(b) represents horizontal bidirectional scanning, 3(c) represents horizontal and vertical bidirectional scanning, and 3(d) represents omnidirectional (horizontal, vertical, diagonal, and anti-diagonal) bidirectional scanning.
[0022] Figure 4 This is a flowchart illustrating the omnidirectional cross-scanning mechanism.
[0023] Figure 5 Images of the disease were collected on-site.
[0024] Figure 6 This is the original image and the label image.
[0025] Figure 7 The diagrams show the disease results inferred by each model, where 7(a) represents Images, 7(b) represents Labels, 7(c) represents PSPNet, 7(d) represents Segmentor, 7(e) represents Segformer, 7(f) represents SegNeXt, and 7(g) represents the present invention.
[0026] Figure 8 The graph shows the trend of FLOPs for each model.
[0027] Figure 9 The graph shows the robustness comparison of each model; the red line represents the model proposed in this invention, the blue line represents Segformer, the orange line represents SegNeXt, the green line represents PSPNet, and the brown line represents Segmentor. The X-axis represents the noise level, and the Y-axis represents mIoU. Detailed Implementation
[0028] The following are specific embodiments of the present invention. It should be noted that the present invention is not limited to the following specific embodiments. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of the present invention.
[0029] Example This embodiment discloses a method for detecting concrete defects based on an omnidirectional cross-scanning mechanism, specifically including the following steps: Step 1: Obtain image datasets of apparent defects in bridge pier concrete and preprocess them; the preprocessing in this embodiment includes data augmentation using geometric distortion and photometric distortion methods.
[0030] In this embodiment, a circular vision acquisition system based on a climbing robot was used to acquire images of apparent defects in the concrete of the Shouchun Bridge pier in Anhui Province, China. These images were then used for model training and analysis. Figure 5 As shown in (a), a host computer-based image visualization software was designed using Microsoft Visual Studio 2019. Images captured by the industrial camera are displayed in real-time and stored on the industrial computer through this software, such as... Figure 5 As shown in (b), 5226 disease images with a resolution of 648×512 were randomly cropped from the acquired original images for model training and validation. In order to prevent overfitting, data augmentation techniques were used, including random photometric distortion, random flipping, and random rotation.
[0031] Step 2: Label the preprocessed image dataset to obtain the labeled image dataset. Randomly divide the labeled image dataset into training set and test set according to the proportion. The labeled image dataset includes segmentation information of multiple bridge pier surface defects. The segmentation information of bridge pier surface defects includes background images, crack images, exposed rebar images and peeling images. This embodiment uses the pixel-level image semantic segmentation and annotation tool Labelme to annotate three disease categories in the disease image: "cracks," "exposed reinforcement," and "stripping." All other interfering objects and the background are set to black. Figure 6 As shown.
[0032] In this embodiment, the labeled image dataset is randomly divided into a training set and a test set in an 8:2 ratio.
[0033] Step 3: Construct a bridge pier defect segmentation network based on an omnidirectional cross-scanning mechanism.
[0034] The bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism in this embodiment includes an encoder and a decoder connected in sequence. The encoder is used to extract multi-scale feature information of the bridge pier apparent defect image, and performs feature enhancement and extraction through multiple omnidirectional state space modules (OSS Blocks) to output multiple feature maps. The decoder is used to perform feature fusion and enhancement on the multiple feature maps output by the encoder through a feature fusion network, and simultaneously uses a fully connected layer to segment the processed feature maps and output the segmentation result.
[0035] The encoder in this embodiment includes four sequentially connected stages. The first stage includes a connected patch embedding module and two OSS Block modules. The second stage includes a connected patch merging module and two OSS Block modules. The third stage includes a connected patch merging module and eight OSS Block modules. The fourth stage includes a connected patch merging module and two OSS Block modules. The omnidirectional state space module OSS Block structure of this embodiment is as follows: Figure 2 As shown in (a).
[0036] The OSS Block module comprises a first part and a second part connected in series. The first part includes a first-layer normalization module (LN) and an omnidirectional selection scanning module (OSSM) connected in sequence, with the output of the OSSM residually connected to the input of the first-layer normalization module (LN). The second part includes a second-layer normalization module (LN) and a feedforward neural network (FFN) module connected in series, with the output of the FFN module also residually connected to the input of the second-layer normalization module (LN). The OSSM module comprises a multilayer perceptron module (MLP), a depthwise separable convolutional model (DW Conv), a SiLU activation function, an omnidirectional selection scanning mechanism (OSS) module, a layer normalization module (LN), and a multilayer perceptron module (MLP) connected in series. The structure of the omnidirectional selection scanning module (OSS) in this embodiment is as follows: Figure 2 As shown in (b).
[0037] The omnidirectional selection scanning mechanism OSS module includes a cascaded scan extension operation module, an SSM module, and a scan merging module, such as... Figure 4 As shown, the scan extension operation module is used to perform cross-scanning of the input image in eight directions: horizontal, vertical, diagonal, and anti-diagonal, generating eight different image block sequences. The scan merging module sums the eight sequences output by the SSM module, merges the global modeling information in multiple directions, and restores the output image to the same shape as the input.
[0038] The omnidirectional cross-scanning mechanism (OSS) structure in this embodiment is as follows: Figure 2 As shown in (c); a schematic diagram of the scanning method in this embodiment is shown below. Figure 3As shown, several commonly used scanning methods include horizontal forward selective scanning (... Figure 3 (a)), Horizontal bidirectional selective scanning ( Figure 3 (b) ), Horizontal and vertical bidirectional selective scanning ( Figure 3 (c) and the omnidirectional (horizontal, vertical, diagonal, and anti-diagonal) bidirectional selective scanning proposed in this invention ( Figure 3 (d)).
[0039] In this embodiment, the decoder sequentially passes all feature maps output by the encoder through the MLP unified feature channel, and then obtains the fused feature map through Concat feature fusion. The fused feature map is then adjusted by MLP to adjust the number of channels to the number of disease categories, and finally the final segmented image is generated by upsampling to obtain the apparent disease results of the bridge pier concrete.
[0040] Step four: Using the training set obtained in step two as input and the segmentation information of the apparent defects of the bridge piers as output, train the bridge pier defect segmentation network based on the omnidirectional cross-scanning mechanism obtained in step three to obtain the bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism.
[0041] The algorithm in this embodiment was developed within the MMSegmentation framework, and all experiments were conducted on a single NVIDIA A800 80GB GPU. Images in the dataset were resized to 512×512 and input into the network for training. The batch size was set to 32, the initial learning rate to 1e-3, and the training epochs to 500. A linear warmup strategy was employed for the first 1000 iterations. The AdamW optimizer was used, and the loss function was a weighted sum of Lovasz loss and Focal loss. Four evaluation metrics were used to assess the performance of the proposed algorithm: mean precision (mP), mean recall (mR), mean F1 score (mF1), and mean intersection-over-union ratio (mIoU).
[0042] Step 5: Input the test set obtained in Step 2 into the pier defect segmentation model based on the omnidirectional cross-scanning mechanism obtained in Step 4 for segmentation to obtain the segmented test set, which includes the apparent defect results of multiple pier concrete.
[0043] This embodiment discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the concrete defect detection method based on an omnidirectional cross-scanning mechanism disclosed in this application.
[0044] This embodiment discloses a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, it implements the concrete defect detection method based on the omnidirectional cross-scanning mechanism disclosed in this application.
[0045] Comparative Example The algorithm proposed in this invention was compared with mainstream models in the field of semantic segmentation, including three CNN-based models: FCN, PSPNet, and Deeplabv 3+; three Transformer-based models: SETR, Segmenter, and DEIT; two improved ViT models: SegFormer and Swin transformer; and two improved CNN architecture models: ConvNeXt and SegNeXt. To ensure the validity of the comparison, all networks were trained and tested using the same data augmentation methods, hyperparameter settings, and loss functions. The model performance comparison results are shown in Table 1.
[0046] Table 1. Model Performance Comparison Results
[0047] The algorithm proposed in this invention has the best mR among all models, while its mP is slightly lower than Deeplabv3+ and SegFormer. This is because mP and mR are inversely proportional; when mR is high, mP is generally slightly lower. The mF1 score represents the harmonic mean of precision and recall, taking into account both aspects of the model. mIoU is used to evaluate the spatial accuracy of the model's predictions and is considered the most important metric for semantic segmentation. Except for mP and mR, the proposed algorithm has the best mF1 and mIoU among all compared models, thus exhibiting superior segmentation performance.
[0048] Table 2 Comparison Results of IoU by Category
[0049] The IoU comparison results for each category are shown in Table 2. Except for cracks, the model proposed in this invention outperforms other models in the segmentation performance of the other categories.
[0050] See Figure 7The results represent the inference results of the best-performing model for each class. In the figure, red represents cracks, green represents exposed tendons, yellow represents peeling, and black represents the background. For targets like cracks, which have local details and continuity, convolutional operations can capture their local structure and texture information very well. This makes several CNN-based models more advantageous in segmenting small and complex cracks, thus their IoU is slightly higher than the model proposed in this invention. Conversely, several ViT-based models perform poorly in crack segmentation. This is because traditional ViT models (SETR, Segmenter, and DEIT) segment the input image into fixed-size patches for processing, without gradually extracting features from low to high levels through multiple convolutions like CNNs. The lack of hierarchical features prevents the model from capturing cracks with strong low-level details, resulting in poor crack segmentation performance.
[0051] The experiment also compared the number of floating-point operations (FLOPs) during the inference phase of the model proposed in this invention with other models, and the comparison results are shown in Table 3. The model proposed in this invention, with a moderate number of parameters, only occupies 42.9G of FLOPs during the inference process, similar to the ViT model with a small number of parameters, and significantly less than several CNN segmentation models. This is not only because the proposed model uses a lightweight decoder with very low FLOPs, but also because the omnidirectional selection scanning module OSSM in the encoder has the advantage of linear complexity.
[0052] Table 3. FLOPs of each model during the inference phase
[0053] Noting that OSSM exhibits linear complexity, separate experimental analysis was conducted. Several different types of backbones were applied to images with resolutions ranging from 384×384 to 2048×2048 for inference, and comparative experiments were performed on FLOPs. By evaluating the FLOPs of the model proposed in this invention as the input image size increases, we can verify whether it inherits the linear complexity characteristic. Table 4 shows the FLOPs of each backbone as the input image resolution increases.
[0054] Table 4 shows the FLOPs (OOM refers to memory overflow) of each backbone network as the input image resolution increases.
[0055] See Figure 8 This shows the changing trend of FLOPs for various backbones as the input image resolution increases. From Figure 8As can be observed, the computational complexity of OSSM and VMamba increases linearly with the increase of input image resolution, becoming comparable to CNN-based backbones such as ResNet-50 and ConvNeXt-T. However, the FLOPs of other ViT-based backbones show a dramatic increase with increasing input image resolution. Furthermore, DeiT and Swin-T even experience insufficient GPU computing resources and memory overflow when inferring images with a resolution of 2048×2048. Therefore, this experiment verifies that the model proposed in this invention has linear complexity and can more effectively process high-resolution disease images.
[0056] Robustness is an important metric for evaluating the performance of deep learning models, typically representing the model's ability to maintain accurate inference when the input data changes. In this experiment, nine real-world image corruptions and perturbations, including noise, digitization, and blurring, with severity levels ranging from 1 to 5, were used to evaluate the robustness of the model proposed in this invention. Table 5 shows the robustness test results under different perturbation conditions. As the severity of the perturbation gradually increases, the segmentation performance of the model proposed in this invention decreases to varying degrees. Table 5. Robustness test results
[0057] See Figure 9 The experimental results show the robustness of the proposed model under different perturbation conditions compared to several other segmentation models. It can be seen that when facing motion blur and image compression with high levels of disruption, the proposed model's segmentation performance is slightly lower than Segformer and SegNeXt. However, when facing the other seven types of perturbation, its segmentation performance is superior to other models, demonstrating better robustness. This is because Mamba's selective scanning mechanism has inherent noise modeling and filtering capabilities. By selectively retaining and forgetting input information, the model can effectively filter out noise and retain important feature information. This mechanism allows the proposed model to extract and retain key features more stably when facing most perturbations, resulting in superior robustness.
[0058] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.
[0059] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.
[0060] Furthermore, the various implementation methods disclosed in this solution can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content invented by this disclosure.
Claims
1. A method for detecting concrete defects based on an omnidirectional cross-scanning mechanism, characterized in that, Specifically, the following steps are included: Step 1: Obtain the image dataset of apparent defects in the concrete of the bridge piers and preprocess it; Step 2: Label the preprocessed image dataset with diseases to obtain the labeled image dataset. Then, randomly divide the labeled image dataset into training set and test set according to the proportion. The labeled image dataset includes segmentation information for multiple bridge pier surface defects, which includes background images, crack images, exposed rebar images, and peeling images. Step 3: Construct a bridge pier defect segmentation network based on an omnidirectional cross-scanning mechanism; Step 4: Using the training set obtained in Step 2 as input and the segmentation information of the apparent defects of the bridge piers as output, train the bridge pier defect segmentation network based on the omnidirectional cross-scanning mechanism obtained in Step 3 to obtain the bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism. Step 5: Input the test set obtained in Step 2 into the bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism obtained in Step 4 for segmentation to obtain the segmented test set, which includes the apparent defect results of multiple bridge pier concrete.
2. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 1, characterized in that, The bridge pier defect segmentation model based on the omnidirectional cross-scanning mechanism includes an encoder and a decoder connected in sequence. The encoder is used to extract multi-scale feature information from images of apparent defects in bridge piers. It enhances and extracts features through multiple omnidirectional state space modules (OSS Blocks) and outputs multiple feature maps. The decoder is used to fuse and enhance the features of multiple feature maps output by the encoder through a feature fusion network, and at the same time, it uses a fully connected layer to segment the processed feature maps and output the segmentation results.
3. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 2, characterized in that, The encoder comprises four stages connected in sequence. The first stage includes a connected patch embedding module and two OSS Block modules. The second stage includes a connected patch merging module and two OSS Block modules. The third stage includes a connected patch merging module and eight OSS Block modules. The fourth stage includes a connected patch merging module and two OSS Block modules.
4. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 3, characterized in that, The OSS Block module comprises a first part and a second part connected in series; The first part includes a first-layer normalization module LN and an omnidirectional selective scanning module OSSM connected in sequence. The output of the omnidirectional selective scanning module OSSM is also residually connected to the input of the first-layer normalization module LN. The second part includes a second-layer normalization module LN and a feedforward neural network (FFN) module connected in sequence. The output of the feedforward neural network (FFN) module is also residually connected to the input of the second-layer normalization module LN.
5. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 4, characterized in that, The OSSM module comprises, in series, a multilayer perceptron module (MLP), a depthwise separable convolutional model (DW Conv), a SiLU activation function, an omnidirectional selection scanning mechanism module (OSS), a layer normalization module (LN), and the multilayer perceptron module (MLP).
6. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 5, characterized in that, The omnidirectional selection scanning mechanism OSS module includes a scan extension operation module, an SSM module, and a scan merging module connected in series. The scanning extension operation module is used to perform cross-scanning of the input image in eight directions: horizontal, vertical, diagonal, and anti-diagonal, generating eight different image block sequences. The scanning merging module sums the eight sequences output by the SSM module, merges global modeling information from multiple directions, and restores the output image to the same shape as the input.
7. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in any one of claims 2-6, characterized in that, The decoder sequentially passes all feature maps output by the encoder through the MLP unified feature channel, and then obtains the fused feature map through Concat feature fusion. The fused feature map is then adjusted by MLP to adjust the number of channels to the number of disease categories, and finally the final segmented image is generated by upsampling to obtain the apparent disease results of the bridge pier concrete.
8. The concrete defect detection method based on omnidirectional cross-scanning mechanism as described in claim 7, characterized in that, The preprocessing includes data augmentation using geometric distortion and photometric distortion methods.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the concrete defect detection method based on an omnidirectional cross-scanning mechanism as described in any one of claims 1-8.
10. A computer program product, characterized in that, Includes a computer program / instruction, which, when executed by a processor, implements the concrete defect detection method based on an omnidirectional cross-scanning mechanism as described in any one of claims 1-8.