Image recognition-based grouting material consolidated body porosity evaluation method and system

Abstract

The application provides a grouting material consolidation body porosity evaluation method and system based on image recognition, and belongs to the field of grouting material performance evaluation, and comprises the following steps: selecting a grouting material consolidation body test block, pretreating the test block to obtain a test sample and establishing a three-dimensional model; determining slice parameters, performing slice operation on the three-dimensional model to generate multiple two-dimensional slice images, and pretreating the two-dimensional slice images; performing resolution processing on the pretreated two-dimensional slice images to construct a slice image dataset; inputting images in a training set of the constructed slice image dataset into a convolutional neural network model for training to obtain a trained convolutional neural network model; processing a grouting material consolidation body to be measured to obtain a pretreated two-dimensional image; inputting the pretreated two-dimensional image into the trained convolutional neural network model to obtain a pore recognition result, and obtaining the porosity of the grouting material consolidation body to be measured based on the pore recognition result.

Description

Technical Field

[0001] This invention belongs to the field of grouting material performance evaluation, and particularly relates to a method and system for evaluating the porosity of grouting material solids based on image recognition. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of civil engineering, grouting technology is a commonly used construction method in areas such as foundation reinforcement, tunnel engineering, and bridge repair. It plays a crucial role in improving the stability of engineering structures, enhancing load-bearing capacity, and preventing leakage. As the core element of grouting technology, the performance of grouting materials directly affects the quality and durability of the project.

[0004] The performance of grouting materials encompasses multiple aspects, with the porosity of the consolidated body being a key performance indicator. Lower porosity generally indicates a denser material, potentially leading to higher strength and better impermeability; while higher porosity may negatively impact the material's mechanical properties and durability. Researchers need to accurately evaluate the porosity of grouting materials under different formulations and preparation processes to gain a deeper understanding of the relationship between material properties and structure, thereby developing grouting materials with superior performance.

[0005] Existing conventional methods for detecting material porosity include density methods, photomicrography, and ultrasonic testing. Density methods and photomicrography are mostly destructive testing methods. In density methods, porosity is calculated by measuring the volume and mass of the material, which often requires cutting or crushing the material, damaging the grout and preventing subsequent performance studies or practical applications on the same batch of material. While non-destructive testing methods such as ultrasonic testing avoid material damage, they suffer from low efficiency, slow speed, and limited accuracy in detecting tiny pores within the material. Because ultrasound wavelengths are relatively long, reflection and scattering from tiny pores are weak, making it difficult to accurately detect their presence, thus complicating the accurate assessment of porosity.

[0006] Furthermore, existing non-destructive testing techniques for material porosity are mostly indirect measurements. They calculate porosity indirectly by measuring physical properties related to pores, rather than directly measuring the absolute value of porosity. While some existing technologies utilize images to characterize pore-fracture structures, such as "Research on Pore-Fracture Structure Characterization and Seepage Characteristics of Rock and Soil Materials Based on Digital Images," this approach is not specifically designed for grouting materials. Secondly, this technology focuses on characterizing the microscopic pore structure of rock and soil and predicting permeability coefficients, without proposing a direct method for predicting material porosity, thus failing to achieve accurate porosity assessment. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides an image recognition-based method for evaluating the porosity of grouting material solids. This method acquires the internal structural information of the grouting material solid sample and reconstructs its three-dimensional structure. The three-dimensional structure is then sliced ​​and analyzed. Pores in the three-dimensional structure slices are identified using machine learning image recognition methods. The porosity of the structure is estimated in three-dimensional space, achieving efficient and non-destructive testing of the grouting material porosity. This method enables direct calculation and evaluation of the material's micropores with higher accuracy.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect discloses a method for evaluating the porosity of grouting material consolidation bodies based on image recognition, including:

[0010] Select solidified grouting material test blocks, pretreat the test blocks to prepare test samples and establish a three-dimensional model;

[0011] Determine the slicing parameters, perform slicing operations on the 3D model to generate multiple 2D slice images, and preprocess the 2D slice images;

[0012] Construct a network structure to perform resolution processing on the preprocessed 2D slice images and build a slice image dataset.

[0013] The images in the training set of the constructed slice image dataset are input into the convolutional neural network model for training, and the trained convolutional neural network model is obtained.

[0014] The solidified grouting material to be tested is processed to obtain a pre-processed two-dimensional image;

[0015] The pre-processed two-dimensional image is input into the trained convolutional neural network model to obtain the porosity identification result, and the porosity of the solidified grouting material to be tested is obtained based on the porosity identification result.

[0016] As a further technical solution, after pre-processing the test block to prepare the test sample, the sample is rotated and scanned from multiple angles to collect two-dimensional projection data of the sample from different angles.

[0017] As a further technical solution, the process of establishing a 3D model is as follows:

[0018] Assume an initial 3D image;

[0019] The initial 3D image is projected onto each acquisition angle to obtain the corresponding 2D projected image;

[0020] The calculated two-dimensional projection image is compared with the actual acquired two-dimensional projection data, and the difference between the two is calculated.

[0021] The 3D image is corrected based on the differences, and the value of each voxel in the 3D image is adjusted to minimize the projection differences.

[0022] Repeat the above steps until a certain convergence condition is met to obtain a three-dimensional model.

[0023] As a further technical solution, determining the slicing parameters includes determining the slice thickness and multiple slicing orientations.

[0024] As a further technical solution, preprocessing of the two-dimensional slice image includes:

[0025] Grayscale adjustment of two-dimensional slice images highlights the difference between pores and surrounding material;

[0026] Enhance the brightness or darkness of the porous area to create a more striking contrast with the surrounding material;

[0027] A denoising algorithm is used to effectively remove noise, improve the signal-to-noise ratio of the image, and reduce interference factors.

[0028] As a further technical solution, a network structure is constructed to perform resolution processing on the preprocessed two-dimensional slice image. The network structure includes a feature extraction layer, a sub-pixel convolutional layer, and an upsampling layer.

[0029] The preprocessed two-dimensional slice image is input into the feature extraction layer, which is usually composed of several convolutional layers and is used to extract the features of the image.

[0030] The image after feature extraction enters the sub-pixel convolutional layer, which learns the sub-pixel displacement information of pixels through convolution operations and outputs a high-resolution feature map.

[0031] The high-resolution feature map is converted into a final high-resolution two-dimensional slice image through an upsampling layer.

[0032] As a further technical solution, the convolutional neural network model adopts deformable convolutional kernels to improve the convolutional neural network. The deformable convolutional kernels consist of a traditional convolutional kernel and an offset learning network.

[0033] Calculating the output of the deformable convolution kernel: Assuming the input image is I, the traditional convolution kernel is K, and the offset is Δp, then the output of the deformable convolution kernel is... Where p is the pixel position of the output image, q is the sampling point position of the traditional convolution kernel, and R is the sampling region of the traditional convolution kernel;

[0034] During the calculation process, the sampling point position of the traditional convolution kernel is first adjusted according to the offset output of the offset learning network. Then, the adjusted convolution kernel is convolved with the input image to obtain the output of the deformable convolution kernel.

[0035] As a further technical solution, the porosity of the solidified grouting material to be tested is obtained based on the porosity identification results. The specific process is as follows:

[0036] Calculate the pore area ratio: For each 2D slice image, count the number of pixels in the pore area, then divide by the total number of pixels in the image to obtain the pore area ratio.

[0037] Porosity is calculated using the volume integral method: the pore area ratio of each slice image is multiplied by the slice thickness to obtain the pore volume corresponding to that slice.

[0038] Then, the pore volumes of all slices are summed and divided by the volume of the entire sample to obtain the porosity of the three-dimensional structure.

[0039] Secondly, a porosity evaluation system for grouting material consolidation based on image recognition is disclosed, including:

[0040] The 3D model building module is configured to: select a solidified grouting material specimen, pre-treat the specimen to prepare test samples and build a 3D model;

[0041] The 2D slice image generation module is configured to: determine slicing parameters, perform slicing operations on the 3D model, generate multiple 2D slice images, and preprocess the 2D slice images;

[0042] The slice image dataset construction module is configured to: construct a network structure, perform resolution processing on the preprocessed two-dimensional slice images, and construct a slice image dataset.

[0043] The convolutional neural network model training module is configured to input images from the training set of the constructed slice image dataset into the convolutional neural network model for training, and obtain a trained convolutional neural network model.

[0044] The porosity assessment module for the solidified grouting material under test is configured to process the solidified grouting material under test to obtain a pre-processed two-dimensional image.

[0045] The pre-processed two-dimensional image is input into the trained convolutional neural network model to obtain the porosity identification result, and the porosity of the solidified grouting material to be tested is obtained based on the porosity identification result.

[0046] The above one or more technical solutions have the following beneficial effects:

[0047] The image recognition-based porosity assessment method for grouting materials proposed in this invention employs non-destructive testing technology, eliminating the need to damage specimens and saving significant specimen resources. This reduces testing costs and avoids additional expenses incurred due to specimen damage. In practical applications, porosity assessment of grouting materials can be performed without compromising the integrity of the engineering structure, providing an economical and efficient solution for engineering quality control. Furthermore, it reduces environmental impact, aligning with the requirements of sustainable development.

[0048] The porosity assessment method proposed in this invention scans the sample, achieving high identification accuracy and providing more precise images, thus offering a reliable foundation for machine learning. A high-efficiency sub-pixel convolutional neural network (ESPCN) is employed for pixel processing of the image, resulting in high-resolution images that better present the detailed information of the pores. Deformable convolutional kernels are also used for neural network training, enabling the convolutional neural network to more flexibly adapt to pores of different shapes and sizes, especially for irregularly shaped and varying-sized micropores, better capturing their features and significantly improving the accuracy of pore identification. This high-precision porosity assessment method is crucial for ensuring the quality of grouting materials and the safety of engineering projects, providing accurate parameter data for engineering design and construction.

[0049] This invention significantly improves the speed of porosity detection for grouting materials through scanning and deep learning image recognition technology. The ESPCN super-resolution algorithm features an efficient sub-pixel convolutional structure, enabling rapid processing of large numbers of two-dimensional slice images, thus improving data processing speed and efficiency and saving time in constructing high-quality slice image datasets. The deformable convolutional kernel, by introducing additional branch networks to learn offsets, improves the adaptability and performance of convolutional neural networks without significantly increasing computational complexity, making model training more efficient. For the specific training model structure after the improved deformable convolutional kernel, please refer to the appendix. Figure 2 As shown.

[0050] Compared to traditional testing methods, this significantly saves time and improves testing efficiency. It is crucial for large-scale engineering construction and quality inspection, as it can obtain a large amount of accurate porosity data in a short time, providing timely information for engineering decisions. Rapid testing can also reduce the time cost of engineering construction, accelerate project progress, and improve the overall efficiency of the project.

[0051] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0052] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0053] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0054] Figure 2 This is a diagram of the training model structure after the improvement of deformable convolution kernels. Detailed Implementation

[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0056] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0057] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0058] Word explanation:

[0059] Efficient Sub-Pixel Convolutional Neural Network (ESPCN) is a network of interconnected sub-pixel convolutional neural networks.

[0060] Example 1

[0061] See appendix Figure 1 As shown, this embodiment discloses a method for evaluating the porosity of grouting material consolidation based on image recognition, including:

[0062] Step 1: Select representative solidified specimens of grouting material, pre-treat the samples to prepare PXCT test specimens and establish a three-dimensional model.

[0063] Specifically, a representative portion of the grouting material's consolidated structure is selected as a sample. A comprehensive assessment is made based on factors such as different locations within the consolidated structure and varying construction conditions to ensure the selected sample's representativeness. The selected sample is then cut and processed to meet the size and shape requirements of the PXCT equipment, preparing it for accurate data acquisition.

[0064] The PXCT device emits X-rays to perform multi-angle rotational scanning of the sample, acquiring X-ray projection data of the sample from different angles to comprehensively obtain the internal information of the sample.

[0065] An iterative reconstruction algorithm was used to perform three-dimensional reconstruction on the acquired two-dimensional projection data. This yielded a three-dimensional model that reveals the internal microstructure, providing a detailed structural basis for subsequent analysis.

[0066] Iterative reconstruction algorithms are methods for reconstructing 3D images through continuous iteration and optimization. The specific implementation method is as follows:

[0067] Initialize the 3D image: First, assume an initial 3D image, which can usually be a uniform low-resolution image.

[0068] Projection calculation: Project the current 3D image onto each acquisition angle to obtain the corresponding 2D projected image.

[0069] Comparison and Correction: The calculated 2D projected image is compared with the actual acquired 2D projected data, and the difference between the two is calculated. Then, the 3D image is corrected based on the difference. Optimization algorithms (such as gradient descent) are typically used to adjust the value of each voxel in the 3D image to minimize the projection difference.

[0070] Repeated iteration: Repeat the above steps until a certain convergence condition is met, such as the projection difference being less than a certain threshold or a certain number of iterations being reached.

[0071] Algorithm formula and explanation: Let the collected two-dimensional projection data be P. i (i represents different acquisition angles), the current 3D image is V, then the projection calculation can be expressed as P. i =Projection(V), where Projection represents the projection function. The correction process can be expressed as: Where α is the step size parameter. It is the gradient of the objective function E, which is usually defined as the difference between the actual projected data and the calculated projected data, i.e.

[0072] In this step, PXCT offers high scanning accuracy, enabling the acquisition of fine internal structures within the grouting material consolidation body. The 3D model created after PXCT scanning reveals the connections and spatial layout between the multi-layered structures within the consolidation body, laying a solid foundation for subsequent pore identification. This approach addresses the issues of low accuracy and insufficient clarity in internal structure exploration associated with traditional testing methods.

[0073] Step 2: Determine the slicing parameters, perform slicing operations on the 3D model to generate multiple 2D slice images, and preprocess the 2D slice images.

[0074] Specifically, the slice thickness and multiple slice orientations are determined. The slice thickness is determined based on the specific analytical requirements and sample size, requiring a smaller thickness to ensure detailed visualization of the sample's internal microstructure. Multiple different angles are selected for the slice orientation, such as along the X, Y, and Z coordinate axes, as well as directions at angles to the axes. To ensure computational efficiency while displaying the sample's internal microstructure in as much detail as possible, multiple slice orientations are chosen to observe the sample from different perspectives, avoiding the limitation of a single direction and preventing the omission of important porosity information. This ensures in-depth analysis of the sample's internal structure from various angles. Each generated 2D slice image must display the fine structure at a specific location within the sample.

[0075] Slicing is performed on the 3D model. Based on determined slicing parameters, the reconstructed 3D model is sliced. Specialized image processing software or algorithms can be used to perform the slicing operation. For example, for a model represented by a 3D array, 2D slice images can be extracted according to specific slicing directions and thicknesses, generating multiple 2D slice images to observe the internal structure of the sample from different angles.

[0076] Preprocessing of two-dimensional slice images mainly includes:

[0077] Grayscale adjustment: Grayscale adjustment is performed on the 2D slice image to highlight the difference between pores and the surrounding material. This can be achieved by adjusting the image's grayscale histogram. For example, adjusting the grayscale value of the pore area to a higher or lower value will make the difference between its grayscale value and the surrounding material greater, making the pores more obvious and facilitating subsequent pore identification.

[0078] Algorithm formula and explanation: Let the grayscale value of the original image be I(x,y) and the grayscale value after adjustment be I′(x,y). The grayscale adjustment can be achieved by the linear transformation formula I′(x,y)=aI(x,y)+b, where a and b are adjustment parameters that need to be selected according to the specific situation.

[0079] Contrast enhancement: This involves increasing the brightness or darkness of porous areas to create a more pronounced contrast with the surrounding material. Contrast enhancement algorithms, such as histogram equalization and local contrast enhancement, can be used. For example, histogram equalization adjusts the image's grayscale histogram to make the grayscale distribution more uniform, thereby enhancing contrast. This makes it easier to identify and analyze porous areas in two-dimensional image slices.

[0080] Algorithm formula and explanation: For histogram equalization, let the gray-level histogram of the original image be H(I), and the cumulative distribution function be... The equalized grayscale value is I′=round((C(I)-C min ) / (C max -C min)), where C min and C max These are the minimum and maximum values ​​of the cumulative distribution function, respectively, and L is the total number of gray levels.

[0081] Denoising: Employing appropriate denoising algorithms effectively removes noise, improves the signal-to-noise ratio of the image, and reduces interference factors. Different denoising algorithms can be selected based on the characteristics of the noise and the features of the image. Possible algorithms include mean filtering, median filtering, and Gaussian filtering. Mean filtering replaces the value of each pixel in the image with the average value of its neighboring pixels; median filtering replaces the value of each pixel with the median value of its neighboring pixels; Gaussian filtering achieves denoising by performing a Gaussian convolution on the image.

[0082] Algorithm formula and explanation: For mean filtering, let N(x,y) be the neighborhood of a pixel (x,y) in the image, then the filtered pixel value is... Where |N(x,y)| is the total number of neighboring pixels.

[0083] Step 3: Collect a large number of preprocessed 2D slice images, use a super-resolution algorithm and an efficient sub-pixel convolutional neural network for resolution processing, and construct a slice image dataset.

[0084] Specifically, using super-resolution algorithms and efficient sub-pixel convolutional neural networks for resolution processing includes:

[0085] The ESPCN network structure is constructed. ESPCN mainly consists of a feature extraction layer, a sub-pixel convolutional layer, and an upsampling layer.

[0086] The collected preprocessed two-dimensional slice images are input into the feature extraction layer, which typically consists of several convolutional layers and is used to extract features from the image.

[0087] Algorithm formula and explanation: Assume the input of the first convolutional layer is of size H×W×C. in The image, where H and W are the height and width of the image, respectively, and C... in The number of input channels is given. The output of this convolutional layer is H×W×C. out C out To determine the number of output channels, the convolution operation can be represented as: Where X is the input image, W is the convolution kernel, and Y is the output image.

[0088] The image after feature extraction enters the subpixel convolutional layer, which learns the subpixel displacement information of pixels through convolution operations.

[0089] It should be explained that, compared to traditional convolutional neural networks, efficient subpixel convolutional neural networks have the following advantages: In terms of upsampling, they employ subpixel convolutional layers and PixelShuffle operations instead of traditional interpolation, resulting in better detail recovery; they are more computationally efficient, as subpixel convolution requires less computation and reduces resource consumption; they have stronger feature representation capabilities, utilizing local features to learn complex mappings to recover high-frequency details; they have better integration capabilities, easily integrating into existing structures with good scalability; and their network structure design is more efficient, achieving better super-resolution results with fewer layers and parameters, reducing complexity and training difficulty while improving generalization ability.

[0090] Algorithm formula and explanation: Assume the input of the sub-pixel convolutional layer is of size H×W×C. in The feature map is output as rH×rW×C out The high-resolution feature map, where r is the magnification factor. The output of the sub-pixel convolutional layer can be represented as I. SR =pixelshuffle(W sp *F(I LR )+b sp ), where I SR It is a low-resolution slice image. Pixelshuffle is a sub-pixel arrangement operation that rearranges the pixels in a low-resolution feature map into a high-resolution image. W sp and b sp These are the weights and biases of the subpixel convolutional layer, respectively.

[0091] Finally, the high-resolution feature map is converted into a final high-resolution two-dimensional slice image through an upsampling layer, enabling the two-dimensional slice image to better present the detailed information of the pores.

[0092] Constructing a slice image dataset using high-resolution images, specifically including:

[0093] Image processing and annotation: The high-resolution 2D slice images processed by ESPCN are processed and classified. The images are annotated according to different pore characteristics, such as pore size, shape, and distribution density.

[0094] Data augmentation: To increase the diversity of the dataset and improve the generalization ability of the model, data augmentation operations can be performed on the training set. Common data augmentation methods include random rotation, flipping, scaling, cropping, and adding noise.

[0095] Finally, the slice image dataset was successfully constructed.

[0096] Step 4: Train a deep learning model using the Convolutional Neural Network (CNN) algorithm, improve the deformable convolution kernel, perform pore identification, and calculate the porosity of the test sample in a three-dimensional plane.

[0097] Specifically, the model is trained using the Convolutional Neural Network (CNN) algorithm, and the specific steps include:

[0098] Constructing a convolutional neural network structure: Classic convolutional neural network architectures such as VGGNet and ResNet can be used, and appropriate adjustments can be made according to the task of aperture recognition. The network typically includes convolutional layers, pooling layers, and fully connected layers.

[0099] Images from the training set of the constructed slice image dataset are input into a convolutional neural network for training. During training, the network's weights and biases are continuously adjusted using the backpropagation algorithm, gradually bringing the network's output closer to the realistic aperture annotations. The core of the backpropagation algorithm is to calculate the gradient of the loss function with respect to the network parameters and update the parameters using gradient descent. Assuming the loss function is L and the network parameters are θ, the parameter update formula is: Where α is the learning rate.

[0100] Algorithm Formula and Explanation: The training process of a convolutional neural network involves numerous matrix operations and optimization algorithms. Taking a simple two-layer convolutional neural network as an example, let the input image be X, the first layer convolutional kernel be W1, the bias be b1, and the activation function be f. Then the output of the first layer is Z1 = f(W1*X + b1), where * represents the convolution operation. The output of the second layer can be calculated similarly. The loss function can be the cross-entropy loss function or the mean squared error loss function, etc. For example, for a binary classification problem, the cross-entropy loss function can be expressed as... Where N is the total number of samples, y i It's a real label. These are the labels predicted by the model.

[0101] The training process is monitored using a validation set. When the loss on the validation set no longer decreases, training is stopped and the network model at this point is saved.

[0102] To address the complex shape of pore images in grouting material consolidation bodies and improve the accuracy of pore identification, a deformable convolutional kernel is used to improve the convolutional neural network. Specific steps include:

[0103] Introducing Deformable Convolutional Kernels: Deformable convolutional kernels are introduced into traditional convolutional neural networks. These kernels consist of a traditional convolutional kernel and an offset learning network.

[0104] Calculating the output of the deformable convolution kernel: Assuming the input image is I, the traditional convolution kernel is K, and the offset is Δp, then the output of the deformable convolution kernel is... Where p represents the pixel position of the output image, q represents the sampling point position of the traditional convolution kernel, and R represents the sampling region of the traditional convolution kernel. During the calculation, the sampling point position of the traditional convolution kernel is first adjusted based on the offset output by the offset learning network. Then, the adjusted convolution kernel is convolved with the input image to obtain the output of the deformable convolution kernel.

[0105] Training the offset learning network: The offset learning network is trained simultaneously with the convolutional neural network (CNN). Its input is the same as the CNN's, and its output is the offset at each sampling point. During training, the parameters of both the CNN and the offset learning network are updated simultaneously using backpropagation, allowing the deformable convolutional kernel to better adapt to apertures of different shapes and sizes. The loss function can be the same as that used in traditional CNNs, such as cross-entropy loss or mean squared error loss. However, considering the special characteristics of the deformable convolutional kernel, regularization terms can be added to constrain the range and variation of the offset. For example, L1 or L2 regularization can be used to limit the size of the offset, preventing excessively large offsets from causing network instability.

[0106] The specific steps for conducting pore identification include:

[0107] The 2D image, preprocessed in step two, is then processed using an efficient sub-pixel convolutional neural network (ESPCN) with a super-resolution algorithm. This processed image is then fed into a trained convolutional neural network, and the network output is the porosity identification result, i.e., the probability that each pixel in the image belongs to a porosity. Ultimately, this method automatically identifies and labels porosity regions, quickly and accurately extracting porosity feature parameters.

[0108] Based on the set threshold, pixels with a probability greater than the threshold are marked as pores, resulting in a binary image of the pore region, where white areas represent pores and black areas represent non-pores.

[0109] When calculating the porosity of the sample, the porosity of the slice is obtained by dividing the number of pixels occupied by the pores by the total number of pixels in the slice. The porosity of each slice is calculated, and finally the porosity of the entire sample can be obtained by using the volume integration method.

[0110] Calculating the porosity of the test sample: Calculate the pore area ratio of each slice image, and then calculate the porosity of the three-dimensional structure using the volume integration method. Specific steps include:

[0111] Calculate the pore area ratio: For each two-dimensional slice image, the recognition result is the probability that each pixel belongs to a pore. Based on the recognition result, count the number of pixels in the pore area, and then divide it by the total number of pixels in the image to obtain the pore area ratio.

[0112] Porosity is calculated using the volume integral method: the pore area ratio of each slice is multiplied by the slice thickness to obtain the pore volume corresponding to that slice. Then, the pore volumes of all slices are summed, and the result is divided by the total volume of the sample to obtain the three-dimensional structural porosity.

[0113] Algorithm formula and explanation: Let p be the proportion of the pore area in the i-th slice image. i If the slice thickness is t and the total volume of the sample is V, then the porosity of the three-dimensional structure is... Where n is the total number of slices.

[0114] In this implementation example, an additional branch network is introduced during the training of the neural network to learn the offset of the convolutional kernels. Based on traditional convolutional kernels, the sampling point position of each convolutional kernel is allowed to be adaptively adjusted according to the learned offset. This allows the convolutional kernels to more flexibly adapt to objects of different shapes and sizes in the image. For targets with irregular shapes and varying sizes, such as tiny holes, deformable convolutional kernels can better capture their features.

[0115] Two-dimensional sliced ​​images that have undergone preprocessing such as grayscale adjustment, contrast enhancement, and noise reduction also need to be processed using the super-resolution algorithm, Efficient Subpixel Convolutional Neural Network (ESPCN).

[0116] A more specific example, Sub-example 1: A method for evaluating the porosity of grouting material consolidation based on image recognition, including:

[0117] Step 1: Sample selection and preprocessing, PXCT scanning and creation of a 3D model.

[0118] Sample selection and pretreatment include:

[0119] Samples were selected from the solidified grouting material of a certain water conservancy project. The selected samples were cut into cubes with a side length of approximately 2 cm to meet the detection size requirements of the PXCT equipment.

[0120] PXCT Scanning and 3D Reconstruction

[0121] The samples were scanned using a PXCT device with the X-ray emission intensity set to a moderate level and the scanning resolution at 0.1 mm. Each sample was scanned at multiple rotational angles with 15° intervals, resulting in the acquisition of X-ray projection data from a total of 24 angles.

[0122] 3D reconstruction was performed using an iterative reconstruction algorithm. The 3D image was initialized as a uniform low-resolution image with all voxels having a value of 120 (grayscale range 0-255). A ray-tracing-based projection function was used for projection calculation. Gradient descent was employed for comparison and correction, with a step size of 0.01. After 120 iterations, the projection difference was less than the set threshold of 0.05, resulting in the 3D reconstructed model.

[0123] Step 2: 3D model slicing and image preprocessing.

[0124] Slicing parameter determination and operation, including:

[0125] The slice thickness was set at 0.4 mm, and the slicing direction was selected along the X, Y, and Z coordinate axes, as well as at a 45° angle to the coordinate axes. Slicing was performed on the 3D reconstructed model according to these parameters, generating a total of 28 2D slice images.

[0126] Two-dimensional slice image preprocessing includes:

[0127] Grayscale adjustment: A linear transformation formula is used for grayscale adjustment, with parameters a = 1.5 and b = -20. The grayscale value of the pore area is adjusted to a higher value to make the difference between its grayscale value and that of the surrounding material more obvious.

[0128] Contrast Enhancement: Histogram equalization algorithm is used. The grayscale histogram of the original image is calculated and then equalized according to the formula, enhancing the contrast between the porous area and the surrounding material.

[0129] Noise Removal: The mean filtering algorithm is used to replace the value of each pixel in the image with the average value of its 3×3 neighboring pixels, effectively removing noise from the image.

[0130] Step 3: Collect two-dimensional slice images, process them with high-resolution algorithms, and construct a slice image dataset.

[0131] Super-resolution processing: 600 two-dimensional slice images with known pore features were collected. These images came from previous research projects on similar grouting materials and some preliminary test results from this experiment. An ESPCN network structure was constructed. The feature extraction layer consisted of three convolutional layers. The first convolutional layer took an image of size 64×64×1 (height and width 64 pixels, input channel 1) as input and output a 64×64×32 feature map. The sub-pixel convolutional layer took the 64×64×32 feature map as input and output a high-resolution 128×128×32 feature map (magnification factor 2).

[0132] The collected low-resolution 2D slice images are input into the ESPCN network for processing. In the feature extraction layer, the convolution kernel size is 3×3, and feature extraction is performed according to the convolution formula. The sub-pixel convolutional layer learns the sub-pixel displacement information of pixels through special convolution operations and outputs a high-resolution feature map according to the formula. Finally, the high-resolution feature map is converted into the final high-resolution 2D slice image through an upsampling layer.

[0133] Dataset construction includes:

[0134] The high-resolution 2D slice images processed by ESPCN were organized and classified. Based on pore size, they were divided into two categories: large pores and small pores, and labeled accordingly. Data augmentation operations were performed on the training set, including random rotations of 90°, 180°, and 270°, horizontal and vertical flipping, and the addition of a small amount of Gaussian noise (mean 0, standard deviation 0.01). Finally, a dataset containing 480 training images, 120 validation images, and 120 test images was constructed.

[0135] Step 4: Train the deep learning model using the Convolutional Neural Network (CNN) algorithm and improve it with deformable convolution kernels, identify pores, and calculate the porosity of the test sample.

[0136] The model training process is as follows:

[0137] A ResNet-18-based convolutional neural network (CNN) architecture was constructed and appropriately modified. The last few fully connected layers were removed, and a convolutional layer was added for feature extraction. Images from the training set of the constructed slice image dataset were input into the CNN for training. The cross-entropy loss function was used, and the learning rate was set to 0.001. During training, the network weights and biases were continuously adjusted using the backpropagation algorithm. After 200 epochs of training, the loss on the validation set no longer decreased, at which point training was stopped, and the network model was saved.

[0138] A deformable convolutional kernel is used to improve the convolutional neural network. A deformable convolutional kernel is introduced, consisting of a traditional 3×3 convolutional kernel and an offset learning network. The offset learning network is trained simultaneously with the convolutional neural network, with the same input and the output being the offset at each sampling point. During training, a cross-entropy loss function is used, and an L2 regularization term (with a weight of 0.001) is added to constrain the range and variation of the offset. The parameters of both the convolutional neural network and the offset learning network are updated simultaneously using the backpropagation algorithm.

[0139] Pore ​​identification and porosity calculation

[0140] The preprocessed image from the second step is input into a trained convolutional neural network. The network output is the identification result of the pores, that is, the probability that each pixel in the image belongs to a pore. A threshold of 0.5 is set, and pixels with a probability greater than 0.5 are labeled as pores, resulting in a binary image of the pore region.

[0141] The porosity of each slice image was calculated. For each 2D slice image, the number of pixels in the porosity region was counted and then divided by the total number of pixels in the image. The porosity of the 3D structure was calculated using the volume integral method. The calculated porosity of the grouting material solidified body in 3D was 3.8%.

[0142] A more specific implementation example, Sub-example 2: A method for evaluating the porosity of grouting material consolidation based on image recognition, including:

[0143] Step 1: Sample selection and pretreatment, PXCT scanning and 3D model creation

[0144] The sample was selected from the solidified grouting material of a building foundation grouting project. The selected sample was cut into a cube with a side length of approximately 3 cm to meet the detection size requirements of the PXCT equipment.

[0145] PXCT Scanning and 3D Reconstruction

[0146] The samples were scanned using a PXCT device with a high X-ray emission intensity and a scanning resolution of 0.08 mm. Each sample was scanned at multiple rotational angles with 10° intervals, resulting in the acquisition of X-ray projection data from a total of 36 angles.

[0147] 3D reconstruction was performed using an iterative reconstruction algorithm. The 3D image was initialized as a uniform low-resolution image with all voxels having a value of 100 (grayscale range 0-255). A ray-tracing-based projection function was used in the simulation. Gradient descent was employed for comparison and correction, with a step size of 0.008. After 150 iterations, the projection difference was less than the set threshold of 0.03, resulting in the obtained 3D reconstructed model.

[0148] Step 2: 3D model slicing and image preprocessing

[0149] Slicing Parameter Determination and Operation

[0150] The slice thickness was set at 0.3 mm, and the slicing direction was selected along the X, Y, and Z coordinate axes, as well as at a 30° angle to the coordinate axes. Slicing was performed on the 3D reconstructed model according to these parameters, generating a total of 36 2D slice images.

[0151] 2D slice image preprocessing

[0152] Grayscale adjustment: A linear transformation formula is used for grayscale adjustment, with parameters a = 2 and b = -40. The grayscale value of the pore area is adjusted to a higher value to make the difference between its grayscale value and that of the surrounding material more obvious.

[0153] Contrast Enhancement: Histogram equalization algorithm is used. The grayscale histogram of the original image is calculated and then equalized according to the formula, enhancing the contrast between the porous area and the surrounding material.

[0154] Noise Removal: The median filtering algorithm is used to replace the value of each pixel in the image with the median of its 5×5 neighboring pixels, effectively removing noise from the image.

[0155] Step 3: Collect 2D slice images, process them using high-resolution algorithms, and construct a slice image dataset.

[0156] Super-resolution processing

[0157] Eighty hundred two-dimensional slice images with known porosity features were collected. These images came from various research projects on grouting materials for building foundations and some preliminary test results from this experiment. An ESPCN network structure was constructed. The feature extraction layer consisted of four convolutional layers. The input to the first convolutional layer was an image of size 80×80×1 (height and width 80 pixels, input channel 1), and the output was a feature map of size 80×80×48. The input to the sub-pixel convolutional layer was the feature map of size 80×80×48, and the output was a high-resolution feature map of size 160×160×48 (magnification factor 2).

[0158] The collected low-resolution 2D slice images are input into the ESPCN network for processing. In the feature extraction layer, the convolution kernel size is 3×3, and feature extraction is performed according to the convolution formula. The sub-pixel convolutional layer learns the sub-pixel displacement information of pixels through special convolution operations and outputs a high-resolution feature map according to the formula. Finally, the high-resolution feature map is converted into the final high-resolution 2D slice image through an upsampling layer.

[0159] Dataset Construction

[0160] The high-resolution 2D slice images processed by ESPCN were organized, classified, and labeled. Data augmentation operations were performed on the training set, including random rotations of 90°, 180°, and 270°, horizontal and vertical flipping, and the addition of a small amount of Gaussian noise (mean 0, standard deviation 0.02). Finally, a dataset containing 640 training images, 160 validation images, and 160 test images was constructed.

[0161] Step 4: Train the deep learning model using the Convolutional Neural Network (CNN) algorithm and improve it with deformable convolution kernels, identify pores, and calculate the porosity of the test sample.

[0162] Model training

[0163] A convolutional neural network (CNN) architecture based on VGGNet was constructed and appropriately modified. The last few fully connected layers were removed, and two convolutional layers were added for feature extraction. Images from the training set of the constructed slice image dataset were input into the CNN for training. The cross-entropy loss function was used, and the learning rate was set to 0.0008. During training, the network weights and biases were continuously adjusted using the backpropagation algorithm. After 300 epochs of training, the loss on the validation set no longer decreased, at which point training was stopped, and the network model was saved.

[0164] A deformable convolutional kernel is used to improve the convolutional neural network. A deformable convolutional kernel is introduced, consisting of a traditional 3×3 convolutional kernel and an offset learning network. The offset learning network is trained simultaneously with the convolutional neural network, with the same input and the output being the offset at each sampling point. During training, a cross-entropy loss function is used, and an L2 regularization term (with a weight of 0.001) is added to constrain the range and variation of the offset. The parameters of both the convolutional neural network and the offset learning network are updated simultaneously using the backpropagation algorithm.

[0165] Pore ​​identification and porosity calculation

[0166] The 36 preprocessed images from the second step are input into a trained convolutional neural network. The network output is the identification result of the pores, that is, the probability that each pixel in the image belongs to a pore. A threshold of 0.55 is set, and pixels with a probability greater than 0.55 are labeled as pores, resulting in a binary image of the pore region.

[0167] The porosity of each slice image was calculated. For each 2D slice image, the number of pixels in the porosity region was counted and then divided by the total number of pixels in the image. The porosity of the 3D structure was calculated using the volume integral method. The calculated porosity of the grouting material solidified body in 3D structure was 4.2%.

[0168] Example 2

[0169] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0170] Example 3

[0171] The purpose of this embodiment is to provide a computer-readable storage medium.

[0172] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0173] Example 4

[0174] The purpose of this embodiment is to provide an image recognition-based system for evaluating the porosity of grouting material consolidation bodies, including:

[0175] The 3D model building module is configured to: select a solidified grouting material specimen, pre-treat the specimen to prepare test samples and build a 3D model;

[0176] The 2D slice image generation module is configured to: determine slicing parameters, perform slicing operations on the 3D model, generate multiple 2D slice images, and preprocess the 2D slice images;

[0177] The slice image dataset construction module is configured to: construct a network structure, perform resolution processing on the preprocessed two-dimensional slice images, and construct a slice image dataset.

[0178] The convolutional neural network model training module is configured to input images from the training set of the constructed slice image dataset into the convolutional neural network model for training, and obtain a trained convolutional neural network model.

[0179] The porosity assessment module for the solidified grouting material under test is configured to process the solidified grouting material under test to obtain a pre-processed two-dimensional image.

[0180] The pre-processed two-dimensional image is input into the trained convolutional neural network model to obtain the porosity identification result, and the porosity of the solidified grouting material to be tested is obtained based on the porosity identification result.

[0181] Example 5

[0182] The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments.

[0183] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0184] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0185] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An image recognition-based method for evaluating the porosity of a grouting material consolidated body, characterized by, The application relates to a method for determining the porosity of a grouting material solidified body. The method comprises the following steps: Selecting a grouting material solidified body test block, pretreating the test block to obtain a test sample, and establishing a three-dimensional model; After the test sample is obtained by pretreating the test block, the sample is subjected to multi-angle rotary scanning, and two-dimensional projection data of the sample is collected from different angles; The process of establishing the three-dimensional model is as follows: An initial three-dimensional image is assumed; The current initial three-dimensional image is projected to each collection angle to obtain a corresponding two-dimensional projection image; The calculated two-dimensional projection image is compared with the actually collected two-dimensional projection data, and the difference between the two is calculated; The three-dimensional image is corrected according to the difference, and the value of each voxel in the three-dimensional image is adjusted so that the projection difference is minimized; The above steps are repeated until a certain convergence condition is met, and the three-dimensional model is obtained; The slice parameters are determined, the three-dimensional model is subjected to a slicing operation to generate a plurality of two-dimensional slice images, and the two-dimensional slice images are pretreated; A network structure is constructed, the resolution of the pretreated two-dimensional slice images is processed, and a slice image dataset is constructed; The images in the training set of the constructed slice image dataset are input into a convolutional neural network model for training, and a trained convolutional neural network model is obtained; Computing the output of the deformable convolution kernel: assuming the input image is , the traditional convolution kernel is , and the offset is , the output of the deformable convolution kernel is , where is the pixel position of the output image, is the sampling point position of the traditional convolution kernel, is the sampling region of the traditional convolution kernel; The convolutional neural network model adopts a deformable convolution kernel to improve the convolutional neural network, and the deformable convolution kernel is composed of a traditional convolution kernel and an offset learning network; In the calculation process, first, the sampling point position of the traditional convolution kernel is adjusted according to the offset output by the offset learning network, and then the adjusted convolution kernel is convolved with the input image to obtain the output of the deformable convolution kernel; The two-dimensional image after pretreatment is obtained by processing the grouting material solidified body to be measured; 2. The image recognition-based evaluation method of the porosity of a cemented body of grouting material according to claim 1, characterized in that, The two-dimensional image after pretreatment is input into the trained convolutional neural network model to obtain a recognition result of pores, and the porosity of the grouting material solidified body to be measured is obtained based on the recognition result of the pores. The two-dimensional slice images are pretreated, including: The two-dimensional slice images are subjected to grayscale adjustment to highlight the difference between the pores and the surrounding material; The brightness or darkness of the pore region is enhanced to form a more distinct contrast with the surrounding material; 3. The image recognition-based evaluation method of the porosity of a grouting material solidification body according to claim 1, characterized in that, A denoising algorithm is adopted to effectively remove noise, improve the signal-to-noise ratio of the image, and reduce interference factors. A network structure is constructed to process the resolution of the pretreated two-dimensional slice images, and the network structure comprises a feature extraction layer, a sub-pixel convolution layer and an up-sampling layer; The pretreated two-dimensional slice images are input into the feature extraction layer, which is composed of a plurality of convolution layers and is used for extracting the features of the images; The images after feature extraction enter the sub-pixel convolution layer, which learns the sub-pixel displacement information of the pixels through convolution operation and outputs a high-resolution feature map; 4. The image recognition-based evaluation method of the porosity of a cemented body of grouting material according to claim 1, characterized in that, The high-resolution feature map is converted into a final high-resolution two-dimensional slice image through the up-sampling layer. The porosity of the grouting material solidified body to be measured is obtained based on the recognition result of the pores, and the specific process is as follows: Calculate the pore area ratio: for each two-dimensional slice image, the number of pixels in the pore region is counted, and then the total number of pixels in the image is divided to obtain the pore area ratio; The volume integral method is used to calculate the porosity: the pore area ratio of each slice image is multiplied by the slice thickness to obtain the pore volume corresponding to the slice. The porosities of all the slices are then summed and divided by the volume of the whole sample to obtain the three-dimensional structure porosity.

5. An image recognition-based grouting material solidification body porosity evaluation system using the method according to claim 1, characterized by, The method comprises the following steps: The three-dimensional model establishing module is configured to select a grouting material consolidation body test block, pretreat the test block to obtain a test sample, and establish a three-dimensional model; The two-dimensional slice image generating module is configured to determine slice parameters, perform a slicing operation on the three-dimensional model, generate a plurality of two-dimensional slice images, and pretreat the two-dimensional slice images; The slice image dataset constructing module is configured to construct a network structure, perform resolution processing on the pretreated two-dimensional slice images, and construct a slice image dataset; The convolutional neural network model training module is configured to input images in a training set of the constructed slice image dataset into a convolutional neural network model for training, and obtain a trained convolutional neural network model; The to-be-tested grouting material consolidation body porosity evaluating module is configured to process a to-be-tested grouting material consolidation body to obtain a pretreated two-dimensional image; The pretreated two-dimensional image is input into the trained convolutional neural network model to obtain a pore recognition result, and a to-be-tested grouting material consolidation body porosity is obtained based on the pore recognition result.

6. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method of any one of claims 1 to 4.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor, when executing the program, implements the steps of the method of any one of claims 1 to 4.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the method of any one of claims 1 to 4.

Images