Concrete strength detection method based on image recognition

Abstract

The invention discloses a concrete strength detection method based on image recognition. The method comprises the following steps: generating a corrected concrete image data set; outputting a shared feature map set; obtaining a semantic mask set by adopting a rasterized pyramid decoder and a channel attention mechanism; performing microstructure feature statistics based on the semantic mask set to obtain a microstructure index set; outputting a concrete strength predicted value set through a transfer learning and fine tuning regression mechanism; performing incremental updating on the strength prediction network by using a pseudo label-active learning loop mechanism to obtain an updated strength prediction network; and reasoning the corrected concrete image data set by using the updated strength prediction network to generate a concrete strength visualization result map set. According to the method, the interference of a non-intensity correlation noise region on a segmentation result is effectively suppressed, and the accuracy of a microstructure index is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of concrete strength testing technology, and in particular to a concrete strength testing method based on image recognition. Background Technology

[0002] With the continuous development of smart construction and infrastructure quality monitoring technologies, concrete, as the main material of building structures, has become a key link in ensuring the quality and safety of engineering projects through strength testing. Traditional concrete strength testing methods mainly rely on non-destructive rebound hammer testing, ultrasonic testing, or destructive core sampling. Among them, the rebound hammer testing method is widely used due to its simple operation and low cost. However, it usually requires multiple measurements on the concrete surface on site and relies on the values ​​of the rebound hammer and empirical formulas for strength estimation. It is easily affected by factors such as construction environment, operating habits, measurement angle, and surface roughness of concrete, resulting in unstable prediction results and certain deviations.

[0003] With the advancement of image processing and deep learning technologies, some studies have begun to attempt strength prediction based on the appearance features of concrete images. However, existing image recognition methods generally suffer from the following technical shortcomings: On the one hand, most image detection methods use fixed-scale feature extraction structures, which are difficult to adapt to the multi-scale spatial distribution characteristics of aggregates, pores, and cracks of different sizes in concrete structures, resulting in insufficient feature representation. On the other hand, existing detection networks have weak response capabilities to local textures in concrete images, making it difficult to effectively capture information on key structural features such as aggregate edges and crack propagation paths. Furthermore, in the strength prediction stage, most methods use direct regression network modeling, lacking explicit modeling and interpretable statistics for microstructural features, making it difficult to accurately reflect the complex relationship between the mechanical properties of concrete and its surface structure.

[0004] Regarding the prediction model update and feedback mechanism, existing image recognition methods typically employ static model training, which involves building an initial model based on a limited sample set. This lacks subsequent dynamic optimization and learning capabilities, making it unable to adapt to changes in concrete samples under different batches and environmental conditions, and resulting in insufficient generalization ability for new data. Summary of the Invention

[0005] One objective of this invention is to propose a concrete strength detection method based on image recognition. This invention effectively suppresses the interference of non-strength-related noise regions on the segmentation results and significantly improves the accuracy of microstructure indicators.

[0006] A concrete strength detection method based on image recognition according to an embodiment of the present invention includes:

[0007] The original concrete image dataset is collected, merged, and corrected to generate a corrected concrete image dataset.

[0008] The corrected concrete image dataset is input into the improved YOLOv12 network detection model. The multi-scale deformation attention module and the adaptive anchor box recombination mechanism are used to obtain the coordinate set of candidate strength indicator regions and output a set of shared feature maps.

[0009] The set of candidate intensity indicator region coordinates and the set of shared feature maps are input into the semantic segmentation sub-network, and the set of semantic masks is obtained by using a raster pyramid decoder and a channel attention mechanism.

[0010] Microstructural feature statistics are performed based on the semantic mask set to obtain a set of microstructural indicators;

[0011] The set of microstructure indicators is input into the strength prediction network, and the set of concrete strength prediction values ​​is output through transfer learning and fine-tuning regression mechanism.

[0012] The set of predicted concrete strength values ​​is compared with the set of field rebound reference strength values. The strength prediction network is incrementally updated using a pseudo-label-active learning loop mechanism to obtain the updated strength prediction network.

[0013] An updated strength prediction network is used to infer the corrected concrete image dataset, generating a set of concrete strength visualization results.

[0014] Optionally, the original concrete image dataset is merged and corrected, including:

[0015] The original concrete image data set is acquired, and the shooting illumination parameters and spatial positioning parameters are recorded simultaneously. Illumination self-correction processing and noise suppression processing are performed on the original concrete image data set to generate a corrected concrete image data set.

[0016] Optionally, the construction of the improved YOLOv12 network detection model includes:

[0017] For each corrected concrete image, input the improved YOLOv12 network detection model;

[0018] For each corrected concrete image, a deformation attention module guided by local texture is constructed by combining the local texture features of the concrete structure. The feature sampling position is dynamically adjusted through a spatial offset adaptive update strategy to form a set of deformation-aware feature maps.

[0019]

[0020] in, This represents the pixel value of the deformation-aware feature map at spatial location p of the i-th corrected concrete image after passing through the local texture-guided deformation attention module in the l-th layer. It reflects the local response of the aggregate, mortar, and pore microstructure. K represents the size of the deformation convolution kernel, and w j This represents the learnable weight parameter at the j-th position of the convolution kernel. This represents the original feature map output from the l-th layer of the YOLOv12 backbone network for the i-th corrected concrete image, containing multi-scale concrete surface texture and structural information. p is the spatial location index on the deformation-aware feature map, representing the specific pixel location to be analyzed in the corrected concrete image. This represents the local spatial offset of the i-th corrected concrete image at the j-th convolutional kernel sampling point in the l-th layer. Guided by the gradient of the microstructure of the concrete surface, it adaptively adjusts to enhance the feature representation of boundary positions such as aggregates and pores, and is defined as:

[0021]

[0022] in, The local texture enhancement coefficient for concrete is generated adaptively based on the local gray-level contrast of the l-scale feature map of the concrete image, which distinguishes the saliency of aggregate and mortar boundaries. The image gradient information represents the i-th corrected concrete image I. i At the corresponding sampling point The local gray-level gradient of the location, ε is a very small positive number to prevent the denominator from being zero, and α is a learnable spatial scale coefficient;

[0023] The set of deformation-sensing feature maps is input into the cross-scale fusion module. For the engineering characteristics of concrete aggregate size span exceeding the threshold, a scale-specific adaptive fusion mechanism is introduced. The fused scale-sensing feature map is a weighted superposition of deformation-sensing feature maps at different scales.

[0024] Based on the scale-aware fusion feature map, an adaptive anchor frame recombination mechanism is used to cluster the size and aspect ratio distribution of the true bounding box of the concrete strength indicator area to obtain an adaptive anchor frame size set. A candidate anchor frame set is generated according to the different spatial positions of the adaptive anchor frame size set in the scale-aware fusion feature map. The cross-union ratio distribution probability of each candidate anchor frame with all true strength indicator area boxes is calculated to obtain the cross-union ratio distribution entropy of each candidate anchor frame. The candidate anchor frame set is optimized and filtered based on the cross-union ratio distribution entropy to obtain a candidate strength indicator area coordinate set.

[0025] Repeat the above steps for all concrete images in the corrected concrete image dataset to finally generate a set of coordinates of all candidate strength indicator regions and a set of all shared feature maps.

[0026] Optionally, the construction of the semantic segmentation sub-network includes:

[0027] In the scale-aware fusion feature map, for each candidate region in the set of candidate intensity indicator region coordinates, the spatial sub-region feature tensor of the candidate intensity indicator region in the scale-aware fusion feature map is extracted based on the center coordinates and size parameters of the candidate region.

[0028] For each spatial sub-region feature tensor, dilated convolution operations with various dilation rates are used to obtain multi-scale context feature sets of the spatial sub-region at different spatial sampling scales;

[0029] For each multi-scale context feature set, the ROIAlign operation is used to map it back to the same spatial resolution as the scale-aware fusion feature map. After the multi-scale context feature sets of all candidate regions are remapped, they are concatenated with the scale-aware fusion feature map according to the channel dimension to generate a candidate region guided fusion feature map.

[0030] The candidate region is guided to input the fused feature map into the channel attention module, the channel weighting coefficient of each channel is calculated, and the fused feature map is weighted by the channel weighting coefficient and then input into the classification head for pixel-level classification prediction, and the semantic mask set is output.

[0031] Optionally, the semantic mask set includes four categories: aggregate semantic mask, mortar semantic mask, void semantic mask, and crack semantic mask. Each type of semantic mask is used to identify the distribution area of ​​corresponding strength-critical structural pixels in the concrete image.

[0032] Optionally, the step of performing microstructural feature statistics based on the semantic mask set includes:

[0033] Perform connected component analysis on the aggregate semantic mask to extract the contour regions of all aggregate targets. For each aggregate target, obtain the minimum bounding rectangle boundary of the aggregate target and calculate the equivalent size of the aggregate target. Perform statistical analysis on the equivalent sizes of all aggregate targets to form an aggregate size distribution histogram. Calculate the standard deviation and maximum size based on the aggregate size distribution histogram.

[0034] Morphological erosion and dilation operations are performed on the hole semantic mask to extract the backbone of the pore structure. For each pore region, a graph traversal algorithm is used to count the number of connected paths between adjacent pixels in the pore region to obtain the connectivity of the pore region. The pore connectivity rate of the concrete image is obtained by averaging the connectivity of all pore regions.

[0035] Local gray-level co-occurrence matrix analysis is performed on the gray-level image region corresponding to the mortar semantic mask. The absolute difference of gray-level values ​​of adjacent pixels is calculated in the horizontal, vertical and two diagonal direction windows. The gray-level difference of all directions and all pixel pairs is averaged to obtain the mortar texture consistency index.

[0036] The main path of the crack should be extracted from the semantic mask of the crack. The minimum number of boxes required to cover the main path of the crack under different box sizes should be calculated based on the box dimension method, and the crack fractal dimension should be calculated.

[0037] The standard deviation of aggregate size distribution, together with the maximum size, pore connectivity, mortar texture consistency index, and crack fractal dimension, form a set of microstructure indices.

[0038] Optionally, the step of outputting a set of predicted concrete strength values ​​through transfer learning and fine-tuning regression mechanisms includes:

[0039] The set of microstructure indices for each corrected concrete image is normalized, and each component after normalization forms a normalized structural feature vector.

[0040] The normalized structural feature vector is input into the strength prediction network based on the transfer learning structure, and the output is the predicted value of concrete strength.

[0041] For each corrected concrete image in the corrected concrete image dataset, the normalization of the microstructure index set, feature vector input, and strength prediction are repeatedly performed to obtain the set of predicted concrete strength values ​​for all concrete images.

[0042] Optionally, the intensity prediction network includes:

[0043] Each predicted concrete strength value in the set of predicted concrete strength values ​​is paired with the corresponding field rebound reference strength value in the set of field rebound reference strength values ​​to generate a predicted-reference strength comparison table.

[0044] For each predicted-reference strength comparison table, the predicted concrete strength value and the field rebound reference strength value are calculated. The absolute value of each prediction error is taken to obtain the set of prediction absolute errors.

[0045] For each prediction absolute error, determine whether it is less than or equal to the error threshold. If it is less than or equal to the error threshold, the sample consisting of the predicted concrete strength value and its corresponding normalized structural feature vector is included in the pseudo-label sample set. Samples that do not meet the condition are included in the undetermined sample set.

[0046] For each sample in the undetermined sample set, the absolute prediction error is calculated as an uncertainty measure. All undetermined samples are sorted from largest to smallest according to the uncertainty measure. The top K samples in terms of uncertainty measure are selected to form an active learning query set. For each sample in the active learning query set, its true intensity label is obtained. Each true intensity label and its corresponding normalized structural feature vector are included in the high confidence label set.

[0047] The pseudo-label sample set and the high-confidence labeled set are merged to obtain the incremental training set, and the parameters of the intensity prediction network are updated to obtain the updated intensity prediction network.

[0048] Optionally, the construction of the set of concrete strength visualization results includes:

[0049] Strength grading standards are extracted from historical concrete mechanical property test data, and grading rules for high-strength concrete, medium-strength concrete, and low-strength concrete are defined based on predicted concrete strength values ​​and concrete structural indicators.

[0050] Based on the predicted concrete strength and corresponding microstructure indices of each corrected concrete image, the strength levels of spatial regions are divided according to grading rules. Each strength level is assigned a unique color label, and different strength level regions are color-coded in the image space to generate a concrete strength visualization result map.

[0051] Optionally, the classification rules for high-strength concrete, medium-strength concrete, and low-strength concrete are as follows:

[0052] High-strength concrete: The predicted concrete strength is ≥60MPa or the maximum size of the aggregate distribution is ≥25mm and the pore connectivity is ≤0.1.

[0053] Medium-strength concrete: 30MPa ≤ predicted concrete strength < 60MPa and the maximum size of aggregate distribution is between 10mm and 25mm;

[0054] Low-strength concrete: The predicted concrete strength is <30MPa or the pore connectivity is >0.1.

[0055] The beneficial effects of this invention are:

[0056] This invention employs an improved YOLOv12 network that integrates local texture perception and spatial offset adaptive mechanisms. In the feature extraction stage, a multi-scale deformation attention module guided by local texture is introduced. This module can combine local grayscale contrast information of aggregate edges, mortar interfaces, and pore contours in concrete images. By dynamically adjusting the convolution sampling position, it achieves adaptive updates of spatial position offset, thereby enhancing the model's responsiveness to structural boundary regions while maintaining the integrity of the original texture features.

[0057] This invention introduces a candidate region-guided fusion mechanism and a channel attention enhancement strategy in the semantic segmentation stage. It combines multi-scale context extraction with dilated convolution and the ROIAlign remapping method to perform refined semantic segmentation of key structures such as cracks, aggregates, and pores in concrete images. Furthermore, it enhances the expressive power of important feature channels by weighting the channel dimensions, effectively suppressing the interference of non-intensity-related noise regions on the segmentation results and significantly improving the accuracy of microstructure indicators.

[0058] This invention introduces a closed-loop self-updating mechanism based on pseudo-labels and active learning. By analyzing the difference between the strength prediction results and the on-site rebound reference strength values, it adaptively selects high-confidence pseudo-label samples and uncertain prediction samples. It then uses active learning methods to prioritize labeling high-uncertainty samples, thereby constructing an incremental training set to update the strength prediction network. This significantly improves the model's generalization ability under different concrete batches, component types, and lighting conditions. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a concrete strength detection method based on image recognition proposed in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0062] refer to Figure 1 A concrete strength detection method based on image recognition, comprising:

[0063] The original concrete image dataset is collected, merged, and corrected to generate a corrected concrete image dataset.

[0064] The corrected concrete image dataset is input into the improved YOLOv12 network detection model. The multi-scale deformation attention module and the adaptive anchor box recombination mechanism are used to obtain the coordinate set of candidate strength indicator regions and output a set of shared feature maps.

[0065] The set of candidate intensity indicator region coordinates and the set of shared feature maps are input into the semantic segmentation sub-network, and the set of semantic masks is obtained by using a raster pyramid decoder and a channel attention mechanism.

[0066] Microstructural feature statistics are performed based on the semantic mask set to obtain a set of microstructural indicators;

[0067] The set of microstructure indicators is input into the strength prediction network, and the set of concrete strength prediction values ​​is output through transfer learning and fine-tuning regression mechanism.

[0068] The set of predicted concrete strength values ​​is compared with the set of field rebound reference strength values. The strength prediction network is incrementally updated using a pseudo-label-active learning loop mechanism to obtain the updated strength prediction network.

[0069] An updated strength prediction network is used to infer the corrected concrete image dataset, generating a set of concrete strength visualization results.

[0070] In this embodiment, the original concrete image dataset is collected, merged, and corrected, including:

[0071] The original concrete image data set is acquired, and the shooting illumination parameters and spatial positioning parameters are recorded simultaneously. Illumination self-correction processing and noise suppression processing are performed on the original concrete image data set to generate a corrected concrete image data set.

[0072] In this embodiment, the construction of the YOLOv12 network detection model is improved, including:

[0073] For each corrected concrete image, input the improved YOLOv12 network detection model;

[0074] The corrected concrete image dataset is a collection of images that have undergone illumination self-correction and noise suppression processing. Each corrected concrete image in the corrected concrete image dataset is processed sequentially.

[0075] For each corrected concrete image, a deformation attention module guided by local texture is constructed by combining the local texture features of the concrete structure. The feature sampling position is dynamically adjusted through a spatial offset adaptive update strategy to form a set of deformation-aware feature maps.

[0076]

[0077] in, This represents the pixel value of the deformation-aware feature map at spatial location p of the i-th corrected concrete image after passing through the local texture-guided deformation attention module in the l-th layer. It reflects the local response of the aggregate, mortar, and pore microstructure. K represents the size of the deformation convolution kernel, and w j This represents the learnable weight parameter at the j-th position of the convolution kernel. This represents the original feature map output from the l-th layer of the YOLOv12 backbone network for the i-th corrected concrete image, containing multi-scale concrete surface texture and structural information. p is the spatial location index on the deformation-aware feature map, representing the specific pixel location to be analyzed in the corrected concrete image. This represents the local spatial offset of the i-th corrected concrete image at the j-th convolutional kernel sampling point in the l-th layer. Guided by the gradient of the microstructure of the concrete surface, it adaptively adjusts to enhance the feature representation of boundary positions such as aggregates and pores, and is defined as:

[0078]

[0079] in, The local texture enhancement coefficient for concrete is generated adaptively based on the local gray-level contrast of the l-scale feature map of the concrete image, which distinguishes the saliency of aggregate and mortar boundaries. The image gradient information represents the i-th corrected concrete image I. i At the corresponding sampling point The local gray-level gradient of the location, ε is a very small positive number to prevent the denominator from being zero, and α is a learnable spatial scale coefficient;

[0080] The local texture-guided deformation attention module enhances the saliency of areas such as aggregate and mortar boundaries during the calculation process using the local texture enhancement coefficient of concrete. The spatial offset is determined by the local gray-level contrast and local gray-level gradient of the concrete image. The spatial offset is weighted by a learnable spatial scale coefficient. The local texture enhancement coefficient is a saliency description of the concrete structure boundary at each spatial location of the concrete image. The local gray-level gradient of the concrete image is used to reflect the gray-level change rate of aggregate, mortar and pore features. The learnable spatial scale coefficient is used for scale adaptation under different concrete imaging conditions. Finally, a deformation-aware feature map set is formed, which is a multi-scale feature set that integrates concrete structural details.

[0081] The set of deformation-sensing feature maps is input into the cross-scale fusion module. For the engineering characteristics of concrete aggregate size span exceeding the threshold, a scale-specific adaptive fusion mechanism is introduced. The fused scale-sensing feature map is a weighted superposition of deformation-sensing feature maps at different scales.

[0082] The scale-specific adaptive fusion mechanism is a feature fusion method designed to address the differences in the distribution characteristics of aggregates, mortar, and pore structures at different spatial scales in concrete images. The scale-specific adaptive fusion mechanism statistically analyzes the response intensity of each key concrete structure in the deformation-aware feature maps at different scales, and adaptively assigns scale fusion weights based on the prominence of concrete strength-related regions in the feature maps at each scale. This makes the mechanism sensitive to strength-indicating structures at different scales. During fusion, the feature maps at different scales are weighted and accumulated to form a scale-aware fusion feature map containing global and local strength information, highlighting the spatial details that best reflect the mechanical properties of concrete, and achieving cross-scale consistency and robustness in detection and segmentation tasks.

[0083] The scale fusion weights in the weighted overlay of scale-aware fusion feature maps are adaptively determined based on the concrete key structural response intensity in the deformation-aware feature maps of each scale. The concrete key structural response intensity is the statistical value of the feature activity of the aggregate, void, and mortar strength indication areas. By adjusting the scale fusion weights, the feature scales that best reflect the concrete strength characteristics are highlighted, and the scale-aware fusion feature maps are obtained.

[0084] Based on the scale-aware fusion feature map, an adaptive anchor frame recombination mechanism is used to cluster the size and aspect ratio distribution of the true bounding box of the concrete strength indicator area to obtain an adaptive anchor frame size set. A candidate anchor frame set is generated according to the different spatial positions of the adaptive anchor frame size set in the scale-aware fusion feature map. The cross-union ratio distribution probability of each candidate anchor frame with all true strength indicator area boxes is calculated to obtain the cross-union ratio distribution entropy of each candidate anchor frame. The candidate anchor frame set is optimized and filtered based on the cross-union ratio distribution entropy to obtain a candidate strength indicator area coordinate set.

[0085] The adaptive anchor frame size set is used to characterize the spatial range of the key strength region in different concrete images. Each candidate anchor frame in the candidate anchor frame set is set according to different anchor frame size parameters in the adaptive anchor frame size set. The cross-union ratio distribution entropy is the spatial overlap distribution uncertainty between the candidate anchor frame and all real strength indication regions. During the screening process, candidate anchor frames with cross-union ratio distribution entropy below the threshold are retained to obtain the candidate strength indication region coordinate set. The candidate strength indication region coordinate set is the set of positioning information of aggregate, void and mortar strength indication structures in concrete images.

[0086] Repeat the above steps for all concrete images in the corrected concrete image dataset to finally generate a set of coordinates of all candidate strength indicator regions and a set of all shared feature maps.

[0087] In this embodiment, the construction of the semantic segmentation sub-network includes:

[0088] In the scale-aware fusion feature map, for each candidate region in the set of candidate intensity indicator region coordinates, the spatial sub-region feature tensor of the candidate intensity indicator region in the scale-aware fusion feature map is extracted based on the center coordinates and size parameters of the candidate region.

[0089] The spatial sub-region feature tensor is used to express the texture information of the local strength structure of concrete aggregates, mortar, and pores within the candidate region.

[0090] For each spatial sub-region feature tensor, dilated convolution operations with various dilation rates are used to obtain multi-scale context feature sets of the spatial sub-region at different spatial sampling scales;

[0091] Multi-scale contextual feature sets are used to enhance the response capability of concrete strength structures with aggregate edges, pore clustering, and mortar distribution at different scales. Different void ratios correspond to different sampling intervals, and void ratio parameter sets are used to describe the sampling scale of all void convolutions.

[0092] For each multi-scale context feature set, the ROIAlign operation is used to map it back to the same spatial resolution as the scale-aware fusion feature map. After the multi-scale context feature sets of all candidate regions are remapped, they are concatenated with the scale-aware fusion feature map according to the channel dimension to generate a candidate region guided fusion feature map.

[0093] Candidate region-guided fusion feature maps integrate multi-scale structural information of the entire image with contextual details of all candidate regions.

[0094] The candidate region is guided to input the fused feature map into the channel attention module, the channel weighting coefficient of each channel is calculated, the fused feature map is weighted by the channel weighting coefficient and then input into the classification head for pixel-level classification prediction, and the semantic mask set is output.

[0095] For each corrected concrete image in the corrected concrete image dataset, the processes of candidate intensity indicator region spatial constraint, multi-scale context feature acquisition, channel fusion, and pixel classification are repeated to obtain the complete semantic mask set.

[0096] In this embodiment, the semantic mask set includes four categories: aggregate semantic mask, mortar semantic mask, hole semantic mask, and crack semantic mask. Each type of semantic mask is used to identify the distribution area of ​​the corresponding strength key structure pixels in the concrete image. The channel weighting coefficient reflects the importance of each channel in expressing the key structure of concrete strength.

[0097] Aggregate semantic masks are used to identify aggregate regions in concrete images, mortar semantic masks are used to identify mortar regions, void semantic masks are used to identify void regions, and crack semantic masks are used to identify crack regions. The spatial dimensions of each semantic mask are completely consistent with the corrected concrete image. The pixel values ​​of all semantic masks are binary distributed, with 0 representing non-target regions and 1 representing target regions.

[0098] In this embodiment, microstructural feature statistics are performed based on the semantic mask set, including:

[0099] Perform connected component analysis on the aggregate semantic mask to extract the contour regions of all aggregate targets. For each aggregate target, obtain the minimum bounding rectangle boundary of the aggregate target and calculate the equivalent size of the aggregate target. Perform statistical analysis on the equivalent sizes of all aggregate targets to form an aggregate size distribution histogram. Calculate the standard deviation and maximum size based on the aggregate size distribution histogram.

[0100] The equivalent size is the square root of the sum of the square of the width and the square of the height of the smallest bounding rectangle of the aggregate target. The equivalent size is used to characterize the particle size of the aggregate target. The aggregate size distribution histogram reflects the distribution of particle sizes of all aggregate targets in the concrete image. The standard deviation is used to represent the dispersion of the aggregate size distribution. The maximum size is used to represent the aggregate target with the largest particle size in the concrete image.

[0101] When calculating the standard deviation based on the aggregate size distribution histogram, the equivalent size of all aggregate targets is statistically analyzed. The equivalent size is used as a set of values ​​to calculate the average value of all aggregate equivalent sizes. The difference between the equivalent size of each aggregate and the average value is calculated, and the squared difference is summed and divided by the total number of aggregate targets. The square root of the result is then taken to obtain the standard deviation of the aggregate size distribution. The maximum size is calculated by selecting the largest value among the equivalent sizes of all aggregate targets. The largest value is the maximum size of the aggregate size distribution.

[0102] Morphological erosion and dilation operations are performed on the hole semantic mask to extract the backbone of the pore structure. For each pore region, a graph traversal algorithm is used to count the number of connected paths between adjacent pixels in the pore region to obtain the connectivity of the pore region. The pore connectivity rate of the concrete image is obtained by averaging the connectivity of all pore regions.

[0103] Pore ​​connectivity is used to measure the connectivity of pore regions in concrete, reflecting the spatial permeability of the overall pore structure.

[0104] Local gray-level co-occurrence matrix analysis is performed on the gray-level image region corresponding to the mortar semantic mask. The absolute difference of gray-level values ​​of adjacent pixels is calculated in the horizontal, vertical and two diagonal direction windows. The gray-level difference of all directions and all pixel pairs is averaged to obtain the mortar texture consistency index.

[0105] The mortar texture consistency index is used to reflect the uniformity of grayscale variation in mortar areas and to describe the uniformity of mortar mixing.

[0106] The main path of the crack should be extracted from the semantic mask of the crack. The minimum number of boxes required to cover the main path of the crack under different box sizes should be calculated based on the box dimension method, and the crack fractal dimension should be calculated.

[0107] The fractal dimension of a crack is used to measure the structural complexity of the main path of the crack. A higher fractal dimension value indicates a more complex crack morphology, while a lower fractal dimension value indicates a simpler crack morphology.

[0108] The standard deviation of aggregate size distribution, along with maximum size, pore connectivity, mortar texture consistency index, and crack fractal dimension, form a set of microstructure indices.

[0109] The semantic mask feature statistical process is repeated for each corrected concrete image in the corrected concrete image dataset to obtain the set of microstructure indicators corresponding to all concrete images.

[0110] In this embodiment, a set of predicted concrete strength values ​​is output through transfer learning and fine-tuning regression mechanisms, including:

[0111] The set of microstructure indices for each corrected concrete image is normalized, and each component after normalization forms a normalized structural feature vector.

[0112] Normalization involves subtracting the mean of the indicators from the historical samples from the original value of each indicator and dividing by the standard deviation of the indicators. The normalized structural feature vector is used to eliminate the dimensional influence between different physical quantities, ensuring the consistency of the contributions of each indicator when predicting the strength of concrete images.

[0113] The normalized structural feature vector is input into the strength prediction network based on the transfer learning structure, and the output is the predicted value of concrete strength.

[0114] The strength prediction network consists of an input layer, two fully connected hidden layers, and an output layer. The predicted concrete strength value is a scalar. The strength prediction network converts the normalized structural feature vector into the predicted concrete strength value through a regression mapping relationship.

[0115] For each corrected concrete image in the corrected concrete image dataset, the normalization of the microstructure index set, feature vector input, and strength prediction are repeatedly performed to obtain the set of predicted concrete strength values ​​for all concrete images.

[0116] In this embodiment, updating the intensity prediction network includes:

[0117] Each predicted concrete strength value in the set of predicted concrete strength values ​​is paired with the corresponding field rebound reference strength value in the set of field rebound reference strength values ​​to generate a predicted-reference strength comparison table.

[0118] For each predicted-reference strength comparison table, the predicted concrete strength value and the field rebound reference strength value are calculated. The absolute value of each prediction error is taken to obtain the set of prediction absolute errors.

[0119] The prediction error is the predicted concrete strength value minus the field rebound reference strength value, resulting in a prediction error set. The absolute prediction error is used to measure the magnitude of the deviation of the strength prediction network on the sample.

[0120] For each prediction absolute error, determine whether it is less than or equal to the error threshold. If it is less than or equal to the error threshold, the sample consisting of the predicted concrete strength value and its corresponding normalized structural feature vector is included in the pseudo-label sample set. Samples that do not meet the condition are included in the undetermined sample set.

[0121] The undetermined sample set consists of all samples whose absolute prediction error is greater than the error threshold.

[0122] For each sample in the undetermined sample set, the absolute prediction error is calculated as an uncertainty measure. All undetermined samples are sorted from largest to smallest according to the uncertainty measure. The top K samples in terms of uncertainty measure are selected to form an active learning query set. For each sample in the active learning query set, its true intensity label is obtained. Each true intensity label and its corresponding normalized structural feature vector are included in the high confidence label set.

[0123] The pseudo-label sample set and the high-confidence labeled set are merged to obtain the incremental training set, and the parameters of the intensity prediction network are updated to obtain the updated intensity prediction network.

[0124] The incremental training set is used to incrementally update the strength prediction network. The mean squared error is used as the incremental loss function. The mean squared error is the average of the squared difference between the predicted concrete strength value of each sample in the incremental training set and the final label. The parameters of the strength prediction network are updated according to the mean squared error to obtain the updated strength prediction network.

[0125] In this embodiment, the construction of the set of concrete strength visualization results charts includes:

[0126] Strength grading standards are extracted from historical concrete mechanical property test data, and grading rules for high-strength concrete, medium-strength concrete, and low-strength concrete are defined based on predicted concrete strength values ​​and concrete structural indicators.

[0127] Based on the predicted concrete strength and corresponding microstructure indices of each corrected concrete image, the strength levels of spatial regions are divided according to grading rules. Each strength level is assigned a unique color label, and different strength level regions are color-labeled in the image space to generate a concrete strength visualization result map.

[0128] Output a set of concrete strength visualization results. The set of concrete strength visualization results includes a spatial distribution map of the strength grades of all corrected concrete images. Each spatial location is displayed with a corresponding color based on the strength prediction value and grading standard.

[0129] In this embodiment, the classification rules for high-strength concrete, medium-strength concrete, and low-strength concrete are as follows:

[0130] High-strength concrete: The predicted concrete strength is ≥60MPa or the maximum size of the aggregate distribution is ≥25mm and the pore connectivity is ≤0.1.

[0131] Medium-strength concrete: 30MPa ≤ predicted concrete strength < 60MPa and the maximum size of aggregate distribution is between 10mm and 25mm;

[0132] Low-strength concrete: The predicted concrete strength is <30MPa or the pore connectivity is >0.1.

[0133] Example 1: At the construction site of the underground section of a rail transit extension project in City A, the preliminary quality assessment report of the concrete in the C5-03 standard section was obtained.

[0134] A total of 6 survey areas were set up at the top of section C5-03, with an average of one survey area every 3 meters along the length direction. Each survey area was divided into 1-meter square areas for shooting, and each area was marked with a number in white ink, from GZ-01 to GZ-06.

[0135] During the data acquisition, the imaging device was placed in the center of the test area and the pitch angle was adjusted to 90 degrees perpendicular to the concrete surface. The real-time illuminance was recorded as 870 Lux. To eliminate the influence of temperature and humidity, the attached temperature and humidity recorder was used to calibrate the ambient temperature to 21.3℃ and the relative humidity to 67%. During the shooting process, he noticed that there were local shrinkage cracks in the GZ-04 area, which were shaped like thin, long, curved bands and were expected to affect the local strength assessment.

[0136] All images were acquired, totaling 1298 images. Each image has a resolution of 5488×3648 pixels. The image data was transmitted synchronously to the analysis terminal in the engineering vehicle via Wi-Fi. The system automatically performed illumination equalization and texture denoising correction, which took about 7 minutes.

[0137] The system enters the region detection and semantic annotation stage. Based on the improved YOLOv12 model, the system first detects the boundaries of aggregate concentration areas, mortar continuous areas and void areas in the image. Among them, the cracks in the GZ-04 area are magnified by the attention mechanism, and two main cracks are clearly distinguished. The longest crack is 28cm and the width is about 0.6mm. The system highlights this area as the main crack area and gives it a structural integrity warning mark.

[0138] Subsequently, the system performed structural feature extraction on all regions, generating microscopic indices, including aggregate size distribution, pore connectivity, mortar texture consistency, and crack fractal dimension. Table 1 below shows some of the actual structural feature sample values:

[0139] Table 1 Sample values ​​of some actual structural features

[0140]

[0141] The indicators in Table 1 are converted into structural feature vectors and then input into the transfer learning prediction network of this invention. Since the model has been pre-trained in the dataset of the National Building Quality Inspection Center (50,000+ concrete images and rebound test comparison data), it only needs to be fine-tuned twice with 80% of the data sampled on site to adapt to the differences in concrete mix proportions and curing on site.

[0142] Table 2 Preliminary predicted values ​​from the model output

[0143]

[0144] The system completes the strength prediction, exports the "C5-03 section concrete image strength prediction report", and simultaneously generates a strength grading visualization map. The GZ-04 area is displayed as an orange-yellow low strength warning area. The comparison results show that the prediction mean square error of the present invention is only 1.16MPa, while that of the traditional rebound method is 3.83MPa. The error is more than 3 times that of the present invention. At GZ-04, the rebound reading is falsely high due to surface cracks, resulting in a serious deviation.

[0145] This invention employs an improved YOLOv12 network that integrates local texture perception and spatial offset adaptive mechanisms. In the feature extraction stage, a multi-scale deformation attention module guided by local texture is introduced. This module can combine local grayscale contrast information of aggregate edges, mortar interfaces, and pore contours in concrete images. By dynamically adjusting the convolution sampling position, it achieves adaptive updates of spatial position offset, thereby enhancing the model's responsiveness to structural boundary regions while maintaining the integrity of the original texture features.

[0146] This invention introduces a candidate region-guided fusion mechanism and a channel attention enhancement strategy in the semantic segmentation stage. It combines multi-scale context extraction with dilated convolution and ROIAlign remapping method to perform refined semantic segmentation of key structures such as cracks, aggregates and pores in concrete images. By weighting the channel dimensions, the expressive power of important feature channels is enhanced, effectively suppressing the interference of non-intensity-related noise regions on the segmentation results and significantly improving the accuracy of microstructure indicators.

[0147] This invention introduces a closed-loop self-updating mechanism based on pseudo-labels and active learning. By analyzing the difference between the strength prediction results and the on-site rebound reference strength values, it adaptively selects high-confidence pseudo-label samples and uncertain prediction samples. It then uses active learning methods to prioritize labeling high-uncertainty samples, thereby constructing an incremental training set to update the strength prediction network. This significantly improves the model's generalization ability under different concrete batches, component types, and lighting conditions.

[0148] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting concrete strength based on image recognition, characterized in that, include: The original concrete image dataset is collected, merged, and corrected to generate a corrected concrete image dataset. The corrected concrete image dataset is input into the improved YOLOv12 network detection model. The multi-scale deformation attention module and the adaptive anchor box recombination mechanism are used to obtain the coordinate set of candidate strength indicator regions and output a set of shared feature maps. The set of candidate intensity indicator region coordinates and the set of shared feature maps are input into the semantic segmentation sub-network, and the set of semantic masks is obtained by using a raster pyramid decoder and a channel attention mechanism. Microstructural feature statistics are performed based on the semantic mask set to obtain a set of microstructural indicators; The set of microstructure indicators is input into the strength prediction network, and the set of concrete strength prediction values ​​is output through transfer learning and fine-tuning regression mechanism. The set of predicted concrete strength values ​​is compared with the set of field rebound reference strength values. The strength prediction network is incrementally updated using a pseudo-label-active learning loop mechanism to obtain the updated strength prediction network. An updated strength prediction network is used to infer the corrected concrete image dataset, generating a set of concrete strength visualization results.

2. The concrete strength detection method based on image recognition according to claim 1, characterized in that, The original concrete image dataset is merged and corrected, including: The original concrete image data set is acquired, and the shooting illumination parameters and spatial positioning parameters are recorded simultaneously. Illumination self-correction processing and noise suppression processing are performed on the original concrete image data set to generate a corrected concrete image data set.

3. The concrete strength detection method based on image recognition according to claim 2, characterized in that, The construction of the improved YOLOv12 network detection model includes: For each corrected concrete image, input the improved YOLOv12 network detection model; For each corrected concrete image, a deformation attention module guided by local texture is constructed by combining the local texture features of the concrete structure. The feature sampling position is dynamically adjusted through a spatial offset adaptive update strategy to form a set of deformation-aware feature maps. in, This represents the pixel value of the deformation-aware feature map at spatial location p of the i-th corrected concrete image after passing through the local texture-guided deformation attention module in the l-th layer. It reflects the local response of the aggregate, mortar, and pore microstructure. K represents the size of the deformation convolution kernel, and w j This represents the learnable weight parameter at the j-th position of the convolution kernel. This represents the original feature map output from the l-th layer of the YOLOv12 backbone network for the i-th corrected concrete image, containing multi-scale concrete surface texture and structural information. p is the spatial location index on the deformation-aware feature map, representing the specific pixel location to be analyzed in the corrected concrete image. This represents the local spatial offset of the i-th corrected concrete image at the j-th convolutional kernel sampling point in the l-th layer. Guided by the gradient of the microstructure of the concrete surface, it adaptively adjusts to enhance the feature representation of boundary positions such as aggregates and pores, and is defined as: in, The local texture enhancement coefficient for concrete is generated adaptively based on the local gray-level contrast of the l-scale feature map of the concrete image, which distinguishes the saliency of aggregate and mortar boundaries. The image gradient information represents the i-th corrected concrete image I. i At the corresponding sampling point The local gray-level gradient of the location, ε is a very small positive number to prevent the denominator from being zero, and α is a learnable spatial scale coefficient; The set of deformation-sensing feature maps is input into the cross-scale fusion module. For the engineering characteristics of concrete aggregate size span exceeding the threshold, a scale-specific adaptive fusion mechanism is introduced. The fused scale-sensing feature map is a weighted superposition of deformation-sensing feature maps at different scales. Based on the scale-aware fusion feature map, an adaptive anchor frame recombination mechanism is used to cluster the size and aspect ratio distribution of the true bounding box of the concrete strength indicator area to obtain an adaptive anchor frame size set. A candidate anchor frame set is generated according to the different spatial positions of the adaptive anchor frame size set in the scale-aware fusion feature map. The cross-union ratio distribution probability of each candidate anchor frame with all true strength indicator area boxes is calculated to obtain the cross-union ratio distribution entropy of each candidate anchor frame. The candidate anchor frame set is optimized and filtered based on the cross-union ratio distribution entropy to obtain a candidate strength indicator area coordinate set. Repeat the above steps for all concrete images in the corrected concrete image dataset to finally generate a set of coordinates of all candidate strength indicator regions and a set of all shared feature maps.

4. The concrete strength detection method based on image recognition according to claim 3, characterized in that, The construction of the semantic segmentation subnetwork includes: In the scale-aware fusion feature map, for each candidate region in the set of candidate intensity indicator region coordinates, the spatial sub-region feature tensor of the candidate intensity indicator region in the scale-aware fusion feature map is extracted based on the center coordinates and size parameters of the candidate region. For each spatial sub-region feature tensor, dilated convolution operations with various dilation rates are used to obtain multi-scale context feature sets of the spatial sub-region at different spatial sampling scales; For each multi-scale context feature set, the ROIAlign operation is used to map it back to the same spatial resolution as the scale-aware fusion feature map. After the multi-scale context feature sets of all candidate regions are remapped, they are concatenated with the scale-aware fusion feature map according to the channel dimension to generate a candidate region guided fusion feature map. The candidate region is guided to input the fused feature map into the channel attention module, the channel weighting coefficient of each channel is calculated, and the fused feature map is weighted by the channel weighting coefficient and then input into the classification head for pixel-level classification prediction, and the semantic mask set is output.

5. The concrete strength detection method based on image recognition according to claim 1, characterized in that, The semantic mask set includes four categories: aggregate semantic mask, mortar semantic mask, void semantic mask, and crack semantic mask. Each type of semantic mask is used to identify the distribution area of ​​corresponding strength-critical structural pixels in the concrete image.

6. The concrete strength detection method based on image recognition according to claim 1, characterized in that, The microstructural feature statistics based on the semantic mask set include: Perform connected component analysis on the aggregate semantic mask to extract the contour regions of all aggregate targets. For each aggregate target, obtain the minimum bounding rectangle boundary of the aggregate target and calculate the equivalent size of the aggregate target. Perform statistical analysis on the equivalent sizes of all aggregate targets to form an aggregate size distribution histogram. Calculate the standard deviation and maximum size based on the aggregate size distribution histogram. Morphological erosion and dilation operations are performed on the hole semantic mask to extract the backbone of the pore structure. For each pore region, a graph traversal algorithm is used to count the number of connected paths between adjacent pixels in the pore region to obtain the connectivity of the pore region. The pore connectivity rate of the concrete image is obtained by averaging the connectivity of all pore regions. Local gray-level co-occurrence matrix analysis is performed on the gray-level image region corresponding to the mortar semantic mask. The absolute difference of gray-level values ​​of adjacent pixels is calculated in the horizontal, vertical and two diagonal direction windows. The gray-level difference of all directions and all pixel pairs is averaged to obtain the mortar texture consistency index. The main path of the crack should be extracted from the semantic mask of the crack. The minimum number of boxes required to cover the main path of the crack under different box sizes should be calculated based on the box dimension method, and the crack fractal dimension should be calculated. The standard deviation of aggregate size distribution, together with the maximum size, pore connectivity, mortar texture consistency index, and crack fractal dimension, form a set of microstructure indices.

7. The concrete strength detection method based on image recognition according to claim 1, characterized in that, The set of predicted concrete strength values ​​output through transfer learning and fine-tuning regression mechanisms includes: The set of microstructure indices for each corrected concrete image is normalized, and each component after normalization forms a normalized structural feature vector. The normalized structural feature vector is input into the strength prediction network based on the transfer learning structure, and the output is the predicted value of concrete strength. For each corrected concrete image in the corrected concrete image dataset, the normalization of the microstructure index set, feature vector input, and strength prediction are repeatedly performed to obtain the set of predicted concrete strength values ​​for all concrete images.

8. The concrete strength detection method based on image recognition according to claim 7, characterized in that, The updated intensity prediction network includes: Each predicted concrete strength value in the set of predicted concrete strength values ​​is paired with the corresponding field rebound reference strength value in the set of field rebound reference strength values ​​to generate a predicted-reference strength comparison table. For each predicted-reference strength comparison table, the predicted concrete strength value and the field rebound reference strength value are calculated. The absolute value of each prediction error is taken to obtain the set of prediction absolute errors. For each prediction absolute error, determine whether it is less than or equal to the error threshold. If it is less than or equal to the error threshold, the sample consisting of the predicted concrete strength value and its corresponding normalized structural feature vector is included in the pseudo-label sample set. Samples that do not meet the condition are included in the undetermined sample set. For each sample in the undetermined sample set, the absolute prediction error is calculated as an uncertainty measure. All undetermined samples are sorted from largest to smallest according to the uncertainty measure. The top K samples in terms of uncertainty measure are selected to form an active learning query set. For each sample in the active learning query set, its true intensity label is obtained. Each true intensity label and its corresponding normalized structural feature vector are included in the high confidence label set. The pseudo-label sample set and the high-confidence labeled set are merged to obtain the incremental training set, and the parameters of the intensity prediction network are updated to obtain the updated intensity prediction network.

9. The concrete strength detection method based on image recognition according to claim 8, characterized in that, The construction of the set of concrete strength visualization results includes: Strength grading standards are extracted from historical concrete mechanical property test data, and grading rules for high-strength concrete, medium-strength concrete, and low-strength concrete are defined based on predicted concrete strength values ​​and concrete structural indicators. Based on the predicted concrete strength and corresponding microstructure indices of each corrected concrete image, the strength levels of spatial regions are divided according to grading rules. Each strength level is assigned a unique color label, and different strength level regions are color-coded in the image space to generate a concrete strength visualization result map.

10. The concrete strength detection method based on image recognition according to claim 9, characterized in that, The classification rules for high-strength concrete, medium-strength concrete, and low-strength concrete are as follows: High-strength concrete: The predicted concrete strength is ≥60MPa or the maximum size of the aggregate distribution is ≥25mm and the pore connectivity is ≤0.

1. Medium-strength concrete: 30MPa ≤ predicted concrete strength < 60MPa and the maximum size of aggregate distribution is between 10mm and 25mm; Low-strength concrete: The predicted concrete strength is <30MPa or the pore connectivity is >0.1.

Images