A method and system for quantitatively analyzing apparent air voids of concrete
By employing semantic segmentation and depth estimation methods, the problem of low efficiency in quantifying bubble defects in concrete structures is solved, achieving efficient and automated bubble defect detection, applicable to various scenarios such as laboratories and construction sites.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHIJIAZHUANG CHANGAN YUCAI BUILDING MATERIALS
- Filing Date
- 2026-04-01
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for the quantitative analysis of air bubble defects in concrete structures are inefficient, subjective, and limited by equipment and space constraints, making automated detection impossible in large structures.
A semantic segmentation and depth estimation-based approach is adopted. First and second semantic segmentation models are used to segment and crop concrete surface images. Plane fitting and perspective correction are performed by combining depth information maps to calculate the real-world area of bubbles.
It achieves second-level analysis results on ordinary mobile devices, eliminates size errors caused by shooting angle, improves detection efficiency and consistency, reduces equipment threshold and cost, and supports rapid detection in a variety of scenarios.
Smart Images
Figure CN122115540A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for quantitative analysis of apparent bubble defects in concrete. Background Technology
[0002] Currently, the construction industry mostly uses hardened concrete pore structure analyzers to quantify air bubble defects in concrete structures. This method is limited by the size of the hardened concrete structure, the equipment is inconvenient to move, and it is inefficient. This testing method can only adjust the formula during the experiment, but the actual production process inevitably differs from the experimental conditions. Furthermore, it cannot automatically quantify air bubble defects in large concrete structures. Manual counting is not only time-consuming but also highly subjective.
[0003] A highly efficient and automated system for quantifying apparent air bubble defects in concrete is urgently needed in the industry. If it were portable, it would greatly facilitate quality inspection work for relevant personnel. The automation of this task relies heavily on semantic segmentation capabilities from artificial intelligence.
[0004] Semantic segmentation, a core task in computer vision, aims to achieve fine-grained classification and understanding of images at the pixel level. With the evolution of deep learning technology, semantic segmentation has progressed from early end-to-end architectures based on fully convolutional networks to multi-scale feature extraction modes that integrate advanced modules such as dilated convolutions, feature pyramids, attention mechanisms, and Transformers. Currently, this technology has matured in terms of feature representation depth, spatial context modeling, and multi-level feature fusion, enabling high-precision contour extraction of target objects in complex scenes. Summary of the Invention
[0005] This invention proposes a quantitative analysis method and system for apparent bubble defects in concrete to solve the problems of low efficiency, strong subjectivity and equipment and site limitations of traditional methods.
[0006] On the one hand, to achieve the above objectives, the present invention provides a method for quantitative analysis of apparent air bubble defects in concrete, comprising: Collect images of hardened concrete surface samples and set size references in the images or obtain the true area information of the concrete area; The hardened concrete surface sample image is segmented into concrete regions using the first semantic segmentation model to obtain a concrete region mask. Based on the concrete region mask, the hardened concrete surface sample image is cropped, and the cropped image is input into the second semantic segmentation model and the depth estimation model respectively to obtain the segmentation mask of the bubble and the size reference object and the depth information map of the concrete region. The depth information map is used to perform planar fitting on the concrete surface, and geometric transformation is used to perform perspective correction on the segmentation mask of the bubble and the size reference. Based on the ratio of the pixel area of the corrected size reference mask to the actual size, or based on the ratio of the actual area of the concrete region to the pixel area, calculate the real-world area of the bubble and the equivalent diameter of the bubble.
[0007] Preferably, the first semantic segmentation model is a convolutional neural network based on an encoder-decoder structure, wherein the encoder part uses multi-scale dilated convolution and spatial pyramid pooling modules to extract features, and the decoder part concatenates the image semantic features and multi-scale features and performs convolution calculation to obtain the concrete region mask.
[0008] Preferably, the processing method of the multi-scale dilated convolution is as follows: ; In the formula, Let r be the feature value at position i of the feature map, r represent the dilatancy rate, and k represent the position of the parameters in the convolution kernel. The parameters of the convolution kernel at position k are... The input feature map contains the feature value at position i+rk.
[0009] Preferably, the second semantic segmentation model is a semantic segmentation model based on the Transformer architecture, and the second semantic segmentation model uses PVTv2 as the backbone network to extract multi-level feature maps of the input image.
[0010] Preferably, the second semantic segmentation model further includes a feature pyramid module and a decoder, wherein the feature pyramid module is used to receive and fuse the multi-level feature maps and output multi-resolution fused features; the decoder is used to segment the bubble and the size reference object from the multi-resolution fused features.
[0011] Preferably, the depth estimation model is a dense depth estimation model based on the DenseNet encoder and convolutional decoder structure, wherein the DenseNet encoder is used to encode the input image into a feature vector, and then the feature vector is input into the convolutional decoder, and processed through several stacked bilinear upsampling and convolution to obtain the depth information map.
[0012] Preferably, the plane fitting employs the RANSAC method; The geometric transformation includes: calculating a rotation matrix based on the normal vector of the fitting plane, rotating the three-dimensional points corresponding to each pixel in the mask, and then reprojecting them onto the image plane.
[0013] On the other hand, to achieve the above objectives, the present invention also provides a quantitative analysis system for apparent air bubble defects in concrete, comprising: The image acquisition module is used to acquire sample images of hardened concrete surfaces and set size reference objects in the images or obtain the true area information of the concrete area. The first segmentation module is used to segment the hardened concrete surface sample image into concrete regions using a first semantic segmentation model to obtain a concrete region mask. The second segmentation module is used to crop the hardened concrete surface sample image based on the concrete region mask, and input the cropped image into the second semantic segmentation model to obtain the segmentation mask of the bubble and the size reference object. The depth estimation module is used to generate depth information maps of concrete areas based on the depth estimation model. The perspective correction module is used to perform planar fitting on the concrete surface using the depth information map, and to perform perspective correction on the segmentation mask of the bubble and the size reference object through geometric transformation. The quantization output module is used to calculate the real-world area of the bubble and its equivalent diameter based on the ratio of the pixel area of the corrected size reference mask to the actual size, or the ratio of the actual area of the concrete region to the pixel area.
[0014] Preferably, the system is deployed on a cloud server and supports receiving images and returning structured bubble quantization data via a network interface.
[0015] Compared with the prior art, the present invention has the following advantages and technical effects: (1) This invention uses semantic segmentation and depth estimation algorithms to collect images using ordinary mobile devices and return analysis results in seconds, which greatly improves detection efficiency and consistency, and makes it possible to conduct rapid and flexible spot checks or full inspections in various scenarios such as laboratories and construction sites, reducing the threshold for professional equipment and the cost of use.
[0016] (2) By introducing a monocular depth estimation algorithm, the present invention estimates the three-dimensional information of the concrete surface from a single image, and then corrects the surface to be parallel to the viewing plane through plane fitting and geometric transformation. This step fundamentally eliminates the size error caused by the shooting angle and achieves engineering-level measurement accuracy under non-contact and non-specific viewing conditions. Attached Figure Description
[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1This is a flowchart of a method for quantitative analysis of apparent bubble defects in concrete according to an embodiment of the present invention. Detailed Implementation
[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0020] This embodiment proposes a method for quantitative analysis of apparent air bubble defects in concrete, including: Collect images of hardened concrete surface samples and set size references in the images or obtain the true area information of the concrete area; The hardened concrete surface sample image is segmented into concrete regions using the first semantic segmentation model to obtain a concrete region mask. Based on the concrete region mask, the hardened concrete surface sample image is cropped, and the cropped image is input into the second semantic segmentation model and the depth estimation model respectively to obtain the segmentation mask of the bubble and the size reference object and the depth information map of the concrete region. The depth information map is used to perform planar fitting on the concrete surface, and geometric transformation is used to perform perspective correction on the segmentation mask of the bubble and the size reference. Based on the ratio of the pixel area of the corrected size reference mask to the actual size, or based on the ratio of the actual area of the concrete region to the pixel area, calculate the real-world area of the bubble and the equivalent diameter of the bubble.
[0021] Specifically, firstly, rulers are affixed to the area to be analyzed or the area is directly measured, and hardened concrete surface samples are collected using various mobile devices. These samples include concrete blocks, cast structures, and precast slab surfaces. Surface bubble defects, rulers, and concrete structures are labeled using segmentation annotation. Then, semantic segmentation is used to segment the concrete region, obtaining the concrete region that needs to be quantified. Subsequently, semantic segmentation is used to segment bubbles and rulers, with the selected model focusing on small target segmentation. During training, augmented data for brightness, contrast, and background diversity are added for the corresponding scene, and high-frequency and low-frequency information is incorporated to optimize the model's perception of bubble edges. This step segments the bubble and ruler regions within the concrete area.
[0022] Next, an open-source depth estimation algorithm was chosen. Since many depth estimation algorithms are trained and tested in natural scenes, and their depth estimation performance fully supports subsequent image correction in this technical solution's scenario, the DenseDepth algorithm was selected to support depth estimation capabilities and calculate the dense depth information of the concrete structure. Finally, based on the structural features extracted from the depth map, perspective distortion in the concrete area was corrected. Spatial resampling was used to complete the orthophoto transformation of the concrete surface. Then, based on a scale or the ratio of the concrete structure area to the bubble area, the equivalent diameter of the bubble was calculated to determine its true area, thus quantifying the bubble defect.
[0023] Furthermore, some defects may occur during the annotation process, making it impossible to determine whether they are caused by bubbles. These targets will be labeled as unclear defects, and these areas will not be included in the loss calculation during bubble defect segmentation.
[0024] Furthermore, the first semantic segmentation model is a convolutional neural network based on an encoder-decoder structure. The encoder part uses multi-scale dilated convolution and spatial pyramid pooling modules to extract features, and the decoder part concatenates the image semantic features and multi-scale features and performs convolution calculation to obtain the concrete region mask.
[0025] Specifically, during the concrete region segmentation process, the image undergoes resolution reduction, normalization, and mean removal sequentially. This is because the concrete structure constitutes the main part of the image and occupies a large area. This approach reduces computational load and hardware requirements.
[0026] The first semantic segmentation model uses DeepLabv3+ based on convolutional neural networks, employing an encoder-decoder structure. The hardened concrete surface sample image is first input into the image feature extraction network in the encoder to extract semantic information from each region of the image.
[0027] To capture targets of different sizes, multi-scale dilated convolution is applied after obtaining the semantic information of the image to encode multi-scale contextual information, thus obtaining multi-scale features. In this embodiment, the output stride is defined as the ratio of the spatial resolution of the input image to the final output resolution. In addition, a spatial pyramid pooling module is added, which extracts the features of multi-scale targets on the convolution by applying dilated convolutions with different dilation rates.
[0028] Dilated convolution increases the receptive field without increasing computational cost compared to regular convolution. The calculation method is as follows: ; In the formula, Let r be the feature value at position i of the feature map, r represent the dilatancy rate, and k represent the position of the parameters in the convolution kernel. The parameters of the convolution kernel at position k are... The input feature map contains the feature value at position i+rk.
[0029] After the encoder computation is complete, the image semantic information and multi-scale features are input into the decoder. The image semantic information is enlarged to four times the input feature tensor size using a bilinear interpolation upsampling encoder, and then concatenated with the multi-scale features. Another 1×1 convolution is applied to the low-level features to reduce the number of channels, aligning it with the number of channels in the decoder. After the two feature sets are concatenated, multiple stacked 3×3 convolutions are used for computation, followed by another bilinear upsampling to restore the original image pixel size, resulting in the concrete region mask.
[0030] Furthermore, the second semantic segmentation model is a semantic segmentation model based on the Transformer architecture. The second semantic segmentation model uses PVTv2 as the backbone network to extract multi-level feature maps of the input image.
[0031] Specifically, in the bubble defect and scale segmentation process, the cropped image only undergoes normalization and mean removal. Since bubble sizes vary and are affected by shooting distance, reducing the resolution may lead to the loss of small targets; therefore, the image resolution input to the network is the original image resolution.
[0032] The second semantic segmentation model adopts a Transformer architecture with PVTv2 as the backbone network and integrates a feature pyramid module and a dual-domain (frequency domain + spatial domain) attention mechanism, which is specifically used for small object segmentation.
[0033] During the training phase, the input image undergoes brightness and contrast adjustments, and background diversity enhancement. Wavelet transform is introduced to extract high-frequency edge information to improve the model's sensitivity to bubble contours. During annotation, suspected defects that cannot be definitively identified as bubbles are marked as "unclear," and these pixels are ignored in the loss calculation.
[0034] The model output is a segmentation mask with two channels, corresponding to the bubble region and the scale region, respectively.
[0035] Furthermore, various frequency maps are decomposed using wavelet transform, including low-frequency information. High-frequency details in the horizontal direction High-frequency details in the vertical direction and high-frequency details in the diagonal direction The summary is as follows: ; ; In the formula, It is a low-frequency characteristic. It is a high-frequency feature.
[0036] Enhanced high-frequency features are obtained by combining high-frequency components with low-frequency features by utilizing global and local channel dependencies between different frequency components. Specifically: ; ; ; In the formula, Linear represents a linear layer. and These represent the ReLU and Sigmoid functions, respectively, and GAP represents the global average pooling operation. This represents element-wise multiplication. Local channel weights are based on local feature extraction. This is for extracting global channel weights based on GAP for global context.
[0037] High frequency features The query is updated via a cross-attention layer, using both key and value pairs. : ; In the formula, The query features are updated after wavelet-enhanced attention. The input query features.
[0038] In the spatial domain, the standard cross-attention mechanism aggregates features by calculating the similarity between all pixel pairs. This dense global interaction easily introduces redundant background information. To filter out irrelevant information and enhance attention to defective regions, a prototype-guided cross-attention mechanism is used. The prototype features are obtained by calculating the flattened spatial dimensions and then applying Softmax. ; In the formula, This represents the features after flattening the convolution calculation. For the prototype feature, T is the transpose operation. Obtaining prototype features Then, through prototype features and query features updated with wavelet-enhanced attention Channel dependency optimization The specific calculation is as follows: ; ; ; In the formula, These are the query features after prototype-enhanced cross-attention calculation.
[0039] Finally, the decoder's output query is fed into the query-based segmentation header for mask prediction, i.e., through... The mask features are calculated by performing a 1 / 4 resolution feature map and a 1×1 convolution, and then flattened. This yields a segmentation mask for the bubble defect and the scale, i.e., a segmentation mask for the bubble and the size reference.
[0040] Furthermore, the depth estimation model is a dense depth estimation model based on the DenseNet encoder and convolutional decoder structure. The DenseNet encoder is used to encode the input image into a feature vector, and then the feature vector is input into the convolutional decoder. The feature vector is processed through several stacked bilinear upsampling and convolution to obtain the depth information map.
[0041] Specifically, the depth estimation model uses the DenseDepth open-source model to extract relative depth information. The image is then processed sequentially according to the concrete region segmentation algorithm logic, including resolution reduction, normalization, and mean removal. Because only the depth information of the planes in the entire scene is needed, there is no need to increase the computational load by using the original resolution.
[0042] Further, the concrete area is corrected. First, a concrete mask is used to extract the corresponding area from the depth information map and the bubble ruler segmentation mask. Discontinuous depth values within the area are removed, and interpolation is used to compensate for the removed values.
[0043] Set the pixel coordinates as the X and Y axes, and the relative depth values as the Z axis. Calculate the surface normals within the mask area, converting each pixel and its depth within the mask into a 3D point.
[0044] The RANSAC method is used to fit the plane equation to obtain the normal vector of 3D points.
[0045] Rotate the plane using the normal vector until it is parallel to the Z-axis to obtain the rotation matrix; rotate each 3D point within the rotated mask area; reproject the rotated points back onto the image plane to obtain the corrected bubble and labeled segmentation mask.
[0046] Finally, based on the pixel area ratio of the bubble and the ruler, or the area ratio of the bubble and the concrete structure, and combined with the real-world area of the ruler or the real-world area of the concrete structure, the real-world area of the bubble is converted, and then the equivalent diameter is calculated.
[0047] This embodiment also provides a quantitative analysis system for apparent air bubble defects in concrete, including: The image acquisition module is used to acquire sample images of hardened concrete surfaces and set size reference objects in the images or obtain the true area information of the concrete area. The first segmentation module is used to segment the hardened concrete surface sample image into concrete regions using a first semantic segmentation model to obtain a concrete region mask. The second segmentation module is used to crop the hardened concrete surface sample image based on the concrete region mask, and input the cropped image into the second semantic segmentation model to obtain the segmentation mask of the bubble and the size reference object. The depth estimation module is used to generate depth information maps of concrete areas based on the depth estimation model. The perspective correction module is used to perform planar fitting on the concrete surface using the depth information map, and to perform perspective correction on the segmentation mask of the bubble and the size reference object through geometric transformation. The quantization output module is used to calculate the real-world area of the bubble and its equivalent diameter based on the ratio of the pixel area of the corrected size reference mask to the actual size, or the ratio of the actual area of the concrete region to the pixel area.
[0048] Furthermore, the system is deployed on a cloud server and supports receiving images and returning structured bubble quantization data via a network interface.
[0049] Specifically, the system is deployed in the cloud, allowing embedded and mobile devices to request services for second-level analysis of 4K images. Furthermore, the system can automatically identify concrete areas in uploaded images, estimate the depth information of these areas, and correct the concrete structure to be parallel to the camera's imaging plane, avoiding the "near objects appear larger, far objects appear smaller" problem. The system quantifies the real-world scale of the bubble based on the provided concrete structure area or a attached ruler.
[0050] Furthermore, this embodiment also provides a specific processing flow for a quantitative analysis system of apparent air bubble defects in concrete, such as... Figure 1 ,include: Measure the concrete area or attach a ruler, and use any device to acquire high-resolution image data of the concrete area; Analyze the image data and input it into the concrete region segmentation algorithm to obtain the concrete region mask; If there is no concrete area, exit the analysis process; otherwise, trim the concrete area, input bubble, scale segmentation algorithm and depth estimation algorithm to obtain bubble, scale mask and depth information map of concrete area. Correct the concrete plane parallel to the camera imaging plane based on the depth information map of the concrete area; Calculate the real-world area of the bubble based on a scale or the area of concrete. The equivalent diameter of the bubble is calculated based on its area and statistically analyzed.
[0051] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for quantitative analysis of apparent air bubble defects in concrete, characterized in that, include: Collect images of hardened concrete surface samples and set size references in the images or obtain the true area information of the concrete area; The hardened concrete surface sample image is segmented into concrete regions using the first semantic segmentation model to obtain a concrete region mask. Based on the concrete region mask, the hardened concrete surface sample image is cropped, and the cropped image is input into the second semantic segmentation model and the depth estimation model respectively to obtain the segmentation mask of the bubble and the size reference object and the depth information map of the concrete region. The depth information map is used to perform planar fitting on the concrete surface, and geometric transformation is used to perform perspective correction on the segmentation mask of the bubble and the size reference. Based on the ratio of the pixel area of the corrected size reference mask to the actual size, or based on the ratio of the actual area of the concrete region to the pixel area, calculate the real-world area of the bubble and the equivalent diameter of the bubble.
2. The method for quantitative analysis of apparent air bubble defects in concrete according to claim 1, characterized in that, The first semantic segmentation model is a convolutional neural network based on an encoder-decoder structure. The encoder part uses multi-scale dilated convolution and spatial pyramid pooling modules to extract features, and the decoder part concatenates the image semantic features and multi-scale features and performs convolution calculation to obtain the concrete region mask.
3. The method for quantitative analysis of apparent air bubble defects in concrete according to claim 2, characterized in that, The processing method for the multi-scale dilated convolution is as follows: ； In the formula, Let r be the feature value at position i of the feature map, r represent the dilatancy rate, and k represent the position of the parameters in the convolution kernel. The parameters of the convolution kernel at position k are... The input feature map contains the feature value at position i+rk.
4. The method for quantitative analysis of apparent air bubble defects in concrete according to claim 1, characterized in that, The second semantic segmentation model is a semantic segmentation model based on the Transformer architecture. The second semantic segmentation model uses PVTv2 as the backbone network to extract multi-level feature maps of the input image.
5. The method for quantitative analysis of apparent bubble defects in concrete according to claim 4, characterized in that, The second semantic segmentation model further includes a feature pyramid module and a decoder. The feature pyramid module is used to receive and fuse the multi-level feature maps and output multi-resolution fused features. The decoder is used to segment the bubble and the size reference object from the multi-resolution fused features.
6. The method for quantitative analysis of apparent air bubble defects in concrete according to claim 1, characterized in that, The depth estimation model is a dense depth estimation model based on the DenseNet encoder and convolutional decoder structure. The DenseNet encoder is used to encode the input image into a feature vector, and then the feature vector is input into the convolutional decoder. The feature vector is processed through several stacked bilinear upsampling and convolution to obtain the depth information map.
7. The method for quantitative analysis of apparent air bubble defects in concrete according to claim 1, characterized in that, The plane fitting was performed using the RANSAC method; The geometric transformation includes: calculating a rotation matrix based on the normal vector of the fitting plane, rotating the three-dimensional points corresponding to each pixel in the mask, and then reprojecting them onto the image plane.
8. A concrete apparent bubble defect quantitative analysis system for implementing the concrete apparent bubble defect quantitative analysis method according to any one of claims 1-7, characterized in that, include: The image acquisition module is used to acquire sample images of hardened concrete surfaces and set size reference objects in the images or obtain the true area information of the concrete area. The first segmentation module is used to segment the hardened concrete surface sample image into concrete regions using a first semantic segmentation model to obtain a concrete region mask. The second segmentation module is used to crop the hardened concrete surface sample image based on the concrete region mask, and input the cropped image into the second semantic segmentation model to obtain the segmentation mask of the bubble and the size reference object. The depth estimation module is used to generate depth information maps of concrete areas based on the depth estimation model. The perspective correction module is used to perform planar fitting on the concrete surface using the depth information map, and to perform perspective correction on the segmentation mask of the bubble and the size reference object through geometric transformation. The quantization output module is used to calculate the real-world area of the bubble and its equivalent diameter based on the ratio of the pixel area of the corrected size reference mask to the actual size, or the ratio of the actual area of the concrete region to the pixel area.
9. The system according to claim 8, characterized in that, The system is deployed on a cloud server and supports receiving images and returning structured bubble quantization data via a network interface.