Segmentation system for concrete ct images and batch processing and quality assessment method

Abstract

This invention discloses a segmentation system and batch processing and quality assessment method for concrete CT images, aiming to solve the problems of low segmentation accuracy, poor batch processing efficiency, and lack of objective quality assessment for multiphase components in complex gray-scale overlapping scenes. The method includes: preprocessing using adaptive filtering based on local gradient statistics; constructing a multi-scale feature pyramid network containing encoder-decoder and skip connections to generate initial segmentation; introducing a channel-spatial dual attention mechanism to optimize microcracks and pore boundaries; achieving parallel task allocation under heterogeneous computing resources through a batch processing scheduling engine; establishing a multi-dimensional quality assessment system covering segmentation consistency, topological integrity, and structural fidelity, and feeding the assessment results back to the preprocessing and model parameter adjustment units to form a closed-loop optimization. This enables high-precision, high-efficiency, and fully automated intelligent segmentation and quality control of concrete CT images.

Description

Technical Field

[0001] This application belongs to the field of image processing and non-destructive testing technology, specifically relating to a segmentation system for concrete CT images and a batch processing and quality assessment method. Background Technology

[0002] With the rapid development of non-destructive testing technology for building materials, X-ray computed tomography (CT) has become a core tool for in-depth research into the microstructure and damage mechanisms of concrete. Concrete, as a typical multiphase composite material, contains various components such as aggregates, hardened cement paste, pores, and microcracks. Three-dimensional CT imaging can reveal its evolution under different loading conditions. In the fields of high-strength concrete research and development and durability assessment of major infrastructure projects, accurate analysis of CT images is not only fundamental to materials science research but also a crucial support for finite element simulation and structural service life prediction.

[0003] Semantic segmentation, efficient batch processing, and result quality assessment of concrete CT images constitute key technological directions in modern computational materials science. This technological direction aims to accurately extract the morphological features of each phase from raw high-dimensional grayscale data through intelligent image processing workflows, process massive amounts of test samples using automated pipelines, and introduce multi-dimensional evaluation indicators to verify the efficiency of the segmentation task, thereby providing standardized data support for the quantitative analysis of concrete materials.

[0004] However, due to the complex gray-level overlap characteristics of concrete components and the interference of scattering noise during CT imaging, traditional edge detection and statistical modeling methods struggle to accurately capture minute pores and cracks, often leading to breaks or misjudgments in structural topological connections. Furthermore, the lack of robust parameter adaptation mechanisms means that existing analysis methods suffer from severe reliance on manual intervention when processing multiple batches of images with varying contrast, significantly limiting the throughput efficiency of large-scale engineering inspections. In addition, feedback adjustment mechanisms for segmentation quality are generally lacking, resulting in a lack of closed-loop verification of segmentation accuracy and fidelity to the original structure, casting doubt on the reliability of the analyzed data in subsequent numerical calculations.

[0005] Therefore, a segmentation system and batch processing and quality assessment method for concrete CT images are desired. Summary of the Invention

[0006] The purpose of this invention is to provide a segmentation system and batch processing and quality assessment method for concrete CT images, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a segmentation system and batch processing and quality assessment method for concrete CT images, comprising the following specific steps: Step 1: Preprocess the original concrete CT image to suppress scattering noise and enhance the gray-level contrast between components. Use an adaptive filtering algorithm based on local gradient statistical characteristics to smooth the image voxels while preserving high-frequency detail information at the aggregate and pore boundaries. Step 2: Construct a multi-scale feature pyramid network model, which includes an encoder, a decoder, and a skip connection structure. The encoder extracts deep semantic features by downsampling layer by layer, the decoder gradually restores spatial resolution by upsampling, and uses skip connections to fuse shallow geometric details of the corresponding levels, thereby generating preliminary semantic segmentation results. Step 3: Introduce an attention-guided mechanism to optimize segmentation boundary accuracy. After upsampling at each stage of the decoder, embed a channel-space dual attention module to dynamically adjust the feature response weights of different regions and enhance the ability to focus on low-contrast targets such as microcracks and small pores. Step 4: Establish a batch processing scheduling engine to automatically allocate computing resources based on the size, resolution, and grayscale distribution characteristics of the input image sequence, execute multiple segmentation tasks in parallel, and record intermediate state parameters during each batch processing. Step 5: Design a multi-dimensional quality assessment system to quantify the effectiveness of the segmentation results from three aspects: segmentation consistency, topological integrity, and structural fidelity. Segmentation consistency is measured by the pixel-level intersection-union ratio index, topological integrity is determined by the Euler number change rate to determine the breakage of connected components, and structural fidelity is reflected by the skeleton matching error to reflect the degree of original morphology restoration. Step 6: Feed the quality assessment results back to the preprocessing and segmentation model parameter adjustment unit to form a closed-loop optimization process. Based on the quality score of the current batch, dynamically adjust the filtering intensity, attention weight distribution and network learning rate to improve the segmentation robustness of subsequent batches.

[0008] Preferably, the adaptive filtering algorithm used in step 1 sets a local smoothing coefficient based on the standard deviation of the gray-level gradient in the neighborhood of each voxel. When the standard deviation is lower than a preset threshold, a stronger smoothing effect is applied to suppress noise, and when the standard deviation is higher than the preset threshold, the smoothing effect is weakened to preserve edge sharpness.

[0009] Preferably, in step 2, the multi-scale feature pyramid network model uses depthwise separable convolution to replace traditional convolution operations in the encoder part, which reduces the number of model parameters while maintaining feature representation ability, and introduces a residual correction module at the end of the decoder to compensate for the spatial information loss caused by multiple upsampling.

[0010] Preferably, in step 3, the channel-spatial dual attention module first aggregates global context information along the channel dimension to generate a channel attention map, then combines the spatial saliency within the local window to calculate a spatial attention map, and finally multiplies the two and applies them to the current feature map to achieve enhanced focus on key regions.

[0011] Preferably, the batch processing scheduling engine in step 4 supports heterogeneous computing architecture, and can dynamically divide task queues according to hardware resource configuration, and coordinately allocate data loading, model inference and result storage operations between the central processing unit and the graphics processing unit to ensure maximum resource utilization under high throughput.

[0012] Preferably, the segmentation consistency index in step 5 not only includes the overall crossover ratio, but also further subdivides the crossover ratio values ​​of the three types of targets: aggregate, slurry, and pores, to identify the segmentation deviation of specific components; the topological integrity assessment determines whether there is over-segmentation or merging by tracking the change in the number of connected components of each phase material before and after segmentation; and the structural fidelity is quantified by extracting the centerline skeleton from the original CT image and the segmentation mask respectively, and calculating the average offset between the two within a predetermined distance range.

[0013] Preferably, the closed-loop optimization process in step 6 includes a quality threshold determination mechanism. When the overall quality score of a certain batch is lower than the preset threshold, the model fine-tuning subroutine is triggered to update only the parameters of several layers in the network that are most relevant to the current quality problem, thus avoiding the computational overhead caused by retraining the entire model.

[0014] The beneficial effects of this invention are: This invention constructs a complete intelligent analysis system for concrete CT images by integrating five core components: adaptive preprocessing, multi-scale semantic segmentation, attention-guided optimization, efficient batch processing scheduling, and multi-dimensional quality assessment. This system can achieve high-precision segmentation of multiphase components in complex gray-scale overlapping scenes without manual intervention, particularly improving the identification capability of micron-level pores and microcracks. Simultaneously, by leveraging a batch processing scheduling engine and a closed-loop feedback mechanism, it significantly improves the processing efficiency and result stability of large-scale image sequences, solving the problems of large performance fluctuations and reliance on expert experience in cross-batch applications of traditional methods. Furthermore, the proposed multi-dimensional quality assessment system not only provides objective criteria for segmentation performance but also provides a reliable data foundation for subsequent finite element modeling and durability analysis, effectively ensuring the integrity and reliability of the information transmission chain from image to engineering decision-making. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall technical solution architecture of the segmentation system and batch processing and quality assessment method for concrete CT images according to an embodiment of this application; Figure 2 This is a schematic diagram of the core principle framework of image segmentation based on multi-scale feature pyramid and channel-space dual attention in the segmentation system and batch processing and quality assessment method for concrete CT images according to the embodiments of this application. Figure 3 This is a flowchart illustrating the closed-loop optimization logic of the multi-dimensional quality assessment system and parameter feedback adjustment in the segmentation system and batch processing and quality assessment method for concrete CT images according to embodiments of this application. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0017] like Figure 1-3 As shown: In this embodiment, a segmentation system and batch processing and quality assessment method for concrete CT images are described in strict accordance with the following steps.

[0018] In the above method, step 1 preprocesses the original concrete CT image to suppress scattering noise and enhance the gray-level contrast between components. An adaptive filtering algorithm based on local gradient statistical characteristics is used to smooth the image voxels while preserving high-frequency details at aggregate and pore boundaries. Specifically, the original concrete CT image is input into the system as a three-dimensional voxel array, with each voxel carrying a gray value reflecting the material density. Due to the scattering effect of X-rays penetrating the concrete specimen and the quantum noise of the detector, the original image often contains granular background noise, and the gray-level boundaries are blurred because the density of cement paste and fine aggregate are similar. When executing the adaptive filtering algorithm, the system first defines a three-dimensional sliding window centered on the current target voxel. The size of this window is set to a voxel cube of 3 x 3 x 3 or 5 x 5 x 5, depending on the image resolution. For all voxels within this sliding window, the local mean of their gray values ​​and the standard deviation of their gray-level gradients are calculated. The calculation of the gray-level gradient covers three mutually perpendicular directions in the spatial coordinate system: the horizontal direction, the vertical direction, and the depth direction. For each voxel, calculate the grayscale difference between it and its neighboring voxels in three directions, and take the square root of the sum of the squares of these differences as the gradient magnitude of that point.

[0019] Furthermore, the adaptive filtering algorithm used in step 1 sets a local smoothing coefficient based on the standard deviation of the gray-level gradient within the neighborhood of each voxel. When the calculated standard deviation of the gray-level gradient is lower than a preset global noise threshold, the system determines that the region is a homogeneous material interior, such as the interior of a large-volume aggregate or a continuous slurry region. In this case, a strong smoothing effect is applied, replacing the gray-level value of the target voxel with the weighted average value within its neighborhood to effectively suppress random scattering noise. When the standard deviation of the gray-level gradient is higher than a preset edge detection threshold, the system determines that the region contains a material boundary, such as the contact zone between aggregate and slurry or the edge of a pore. In this case, the smoothing effect is weakened, or even the original gray-level value is kept unchanged, thereby preventing the loss of high-frequency details and boundary blurring caused by the filtering operation. In this way, the preprocessing process can dynamically adjust the filtering intensity according to the complexity of the local structure of the image.

[0020] In the above method, step 2 involves constructing a multi-scale feature pyramid network model. This model includes an encoder, a decoder, and skip connections. The encoder extracts deep semantic features through layer-by-layer downsampling, while the decoder gradually restores spatial resolution through upsampling operations and utilizes skip connections to fuse shallow geometric details from corresponding levels, thereby generating preliminary semantic segmentation results. Specifically, the multi-scale feature pyramid network is built on a deep convolutional neural network, and its input is a 3D voxel block preprocessed in step 1. The encoder part is designed with 5 consecutive downsampling levels. Each level contains two sets of convolutional operations with a kernel size of 3x3x3, and a pooling layer or strided convolutional layer with a stride of 2. This is used to reduce the spatial size of the feature map to half of its original size while doubling the number of feature channels to capture higher-level semantic abstraction information.

[0021] Preferably, in step 2, the multi-scale feature pyramid network model uses depthwise separable convolution instead of traditional convolution operations in the encoder part. Depthwise separable convolution decomposes standard convolution into two stages: channel-wise convolution and pointwise convolution. In the channel-wise convolution stage, the system independently applies a convolution kernel to each channel of the input feature map for spatial feature extraction; in the subsequent pointwise convolution stage, a 1x1x1 convolution kernel is used to linearly combine the outputs of all channels. This design greatly reduces the number of model parameters and computational complexity, making it particularly suitable for processing massive amounts of concrete 3D CT data, while maintaining strong feature representation capabilities by decoupling spatial and channel correlations. The decoder part performs a process symmetrical to the encoder, recovering the image size step by step through bilinear interpolation or transposed convolution operations. In each layer of the decoder, feature maps from the corresponding layers of the encoder are directly stitched into the current feature stream through skip connections. This stitching operation allows the network to determine the material category using deep semantic information while simultaneously tracing back the precise geometric contours in shallow features. In addition, a residual correction module is introduced at the end of the decoder. This module contains two residual branches. By learning the residual value between the prediction result and the real label, it compensates for the loss of spatial positioning information caused by multiple downsampling and upsampling operations.

[0022] In the above method, step 3 introduces an attention-guided mechanism to optimize segmentation boundary accuracy. After upsampling at each level of the decoder, a channel-spatial dual attention module is embedded to dynamically adjust the feature response weights of different regions, enhancing the ability to focus on low-contrast targets such as microcracks and small pores. Specifically, in complex concrete CT images, microcracks often occupy only a few voxels in width and are easily ignored during feature transfer. By embedding a channel-spatial dual attention module at each feature fusion node of the decoder, the system can selectively enhance key features.

[0023] Preferably, the channel-spatial dual attention module in step 3 first aggregates global context information along the channel dimension to generate a channel attention map. Specifically, the input feature map is subjected to global average pooling and global max pooling to obtain two vectors describing the global attributes of the channels. These two vectors are then fed into a multilayer perceptron network with shared weights, and processed by an activation function to obtain channel attention weights. These weights reflect the importance of different feature channels to the current segmentation task; for example, channels specifically for extracting edge features will be given higher weights. Next, the module calculates the spatial attention map by combining spatial saliency within the local window. The system performs average pooling and max pooling operations along the channel axis on the feature map after channel weight adjustment, and concatenates the two resulting two-dimensional spatial maps. A spatial attention mask is then generated through a 7x7 convolutional layer and an activation function. Finally, the channel attention weights are multiplied by the spatial attention mask, and the result is applied to the original feature map to enhance the focus on key material regions (such as microcrack clusters) and weaken the response to background noise regions.

[0024] In the above method, step 4 establishes a batch processing scheduling engine that automatically allocates computing resources based on the size, resolution, and grayscale distribution characteristics of the input image sequence, executes multiple segmentation tasks in parallel, and records intermediate state parameters during each batch processing. In practical engineering applications, CT scan data of concrete specimens typically contains thousands of slices, totaling tens of gigabytes. The batch processing scheduling engine is responsible for scientifically dividing these massive amounts of data into tasks. The engine first scans the input directory to obtain metadata for all sequences to be processed, including the total number of slices, the pixel dimension of a single slice, and the grayscale histogram distribution.

[0025] Preferably, the batch processing scheduling engine in step 4 supports heterogeneous computing architectures and can dynamically divide task queues according to hardware resource configurations. The engine uses a built-in resource monitoring program to monitor the load of the central processing unit (CPU), the video memory usage of the graphics processing unit (GPU), and the available memory space of the system in real time. During processing, the engine adopts a multi-threaded parallel mode, performing data prefetching, format decompression, and the preprocessing operations described in step 1 on the CPU side, while deploying the deep learning inference models described in steps 2 and 3 on the GPU side. To ensure maximum resource utilization under high throughput, the scheduling engine dynamically adjusts the batch size of image slices sent to the GPU based on the current remaining video memory. If sufficient video memory is detected, the batch size is increased to fully utilize the computational parallelism of the GPU; if video memory is close to the warning line, the task is automatically split into smaller slice groups, and multi-stream concurrency technology is activated to achieve overlapping coverage of data transmission and computational inference.

[0026] In the above method, step 5 involves designing a multi-dimensional quality assessment system to quantify the effectiveness of the segmentation results from three aspects: segmentation consistency, topological integrity, and structural fidelity. Specifically, after obtaining the initial segmentation mask, the system does not output it directly; instead, the quality assessment module performs an automated check.

[0027] Preferably, the segmentation consistency index in step 5 not only includes the overall Intersection over Union (IoU), but also further subdivides the IoU values ​​for the three target categories: aggregate, slurry, and pores. The IoU is calculated as follows: for a specific component (such as pores), the volume of the intersection between the set of voxels predicted by the statistical model as pores and the set of voxels manually labeled or labeled as pores by a high-precision benchmark is divided by the volume of their union. By comparing the IoU values ​​of different components, the system can identify segmentation deviations for specific components. For example, when the pore IoU is much lower than the aggregate IoU, it indicates that the model's sensitivity to hollow structures is insufficient. Topological integrity assessment is performed by tracking the change in the number of connected components of each phase before and after segmentation. In specific operation, the system uses a three-dimensional connected component labeling algorithm to count the number of each independent aggregate particle in the segmentation result and calculate the Euler number in topology. If the rate of change of the Euler number after segmentation exceeds a preset ratio, it is determined that there is over-segmentation (i.e., a single aggregate is cut into multiple parts) or merging (i.e., two adjacent aggregates are incorrectly connected). Structural fidelity assessment is achieved by extracting the centerline skeleton from the edge gradient map of the original CT image and the edge map of the segmentation mask, respectively. The system uses a thinning algorithm to extract the skeleton lines of the material and calculates the average offset of the two sets of skeleton lines within a predetermined distance range in three-dimensional space. The smaller the average offset, the higher the accuracy of the segmentation result in restoring the original geometry.

[0028] In the above method, step 6 feeds back the quality assessment results to the preprocessing and segmentation model parameter adjustment unit to form a closed-loop optimization process. Based on the quality score of the current batch, the filtering intensity, attention weight distribution and network learning rate are dynamically adjusted to improve the segmentation robustness of subsequent batches.

[0029] Preferably, the closed-loop optimization process in step 6 includes a quality threshold determination mechanism. When the overall quality score of a batch (calculated by weighting consistency, topology, and fidelity) is lower than a preset minimum qualified threshold, the system determines that the current model parameters do not match the features of the batch of samples. At this time, the model fine-tuning subroutine is triggered. The system does not start retraining the entire model, but only updates the parameters of several layers in the network that are most relevant to the current quality problem. For example, if the feedback shows poor topological integrity and large edge offset, the system will specifically increase the weight of the spatial attention module in step 3 and fine-tune the parameters of the residual correction module in step 2, while appropriately reducing the filtering smoothing coefficient in the preprocessing stage. This local fine-tuning strategy can quickly adapt to the feature differences of different batches of concrete samples (such as different aggregate gradations, background grayscale fluctuations caused by water-cement ratio, etc.), ensuring processing efficiency while avoiding the huge computational overhead and catastrophic forgetting problem caused by retraining the entire model.

[0030] To further illustrate the implementation details of this invention, a specific operational scenario is described below. Assume the object to be processed is a set of CT scan images of a high-strength concrete specimen, totaling 2000 images with a resolution of 1024 x 1024 pixels.

[0031] In step 1, the system loads the first batch of images and calculates that the images contain a large number of fine needle-like aggregates. The adaptive filtering algorithm identifies these needle-like edges by calculating the standard deviation of the gray-level gradient within the voxel neighborhood. At this point, the smoothing coefficient is set to a minimum value (e.g., 0.1), while in the broader cement paste area, the smoothing coefficient is automatically adjusted to 0.8. This dynamic adjustment process is entirely driven by local statistical characteristics and requires no manual intervention.

[0032] In step 2, the preprocessed image is fed into a multi-scale feature pyramid network. Due to the use of depthwise separable convolution, the model maintains high feature extraction accuracy while significantly reducing the time for a single forward inference pass. The skip connection structure continuously feeds high-resolution boundary information from the lower layers to the deeper layers, ensuring the accuracy of aggregate edge positioning.

[0033] In step 3, the channel-space dual attention module plays a crucial role in addressing the micron-level pores that may exist in concrete. By enhancing the filter response sensitive to pore features in the channel dimension and locking in regions of high gradient change in the spatial dimension, the model successfully captures microcracks that are easily missed in ordinary segmentation methods.

[0034] In step 4, the batch scheduling engine detects that the current system has two high-performance GPUs, so it splits the 2000 images into two parallel queues, with each queue handling 1000 images. The engine monitors memory fluctuations in real time to ensure that memory usage remains at 85% when performing complex attention operations, achieving optimal allocation of hardware performance.

[0035] In step 5, the quality assessment system randomly checks the segmentation results of the first batch of 100 images. The calculated consistency index of the aggregate reached over 95%, but the topological integrity of the pores was slightly low, and Euler number analysis showed that some slender pores were fractured.

[0036] In step 6, based on the evaluation feedback from step 5, the system automatically triggers closed-loop adjustment logic. Before processing the next batch of images, the system fine-tunes the residual correction parameters at the decoder end and increases the sensitivity threshold of spatial attention. After adjustment, the porosity of subsequent batches is significantly improved, and the overall segmentation quality score remains stable above the preset high threshold.

[0037] Example 2 This embodiment, based on Embodiment 1, further refines and expands the system and method for complex CT images of concrete containing a large amount of recycled aggregate. Recycled aggregate concrete has a more complex grayscale hierarchy than ordinary concrete due to the presence of an internal layer of old mortar, and multiple interfaces exist between the aggregate and the old mortar, and between the old mortar and the new mortar.

[0038] In step 1, considering the multi-interface characteristics of recycled aggregate concrete, the adaptive filtering algorithm adds a multi-scale window analysis function. The system not only calculates the gradient standard deviation within a 3x3x3 neighborhood, but also simultaneously calculates the macroscopic grayscale distribution within a 7x7x7 range. When a small-scale window displays a high gradient while a large-scale window displays a low gradient change, the system determines the current location as a subtle interface and employs a hybrid smoothing strategy based on a nonlinear combination of median filtering and Gaussian filtering. Specifically, the smoothing coefficient is no longer a single numerical value, but a function that varies with the complexity of local features. This method suppresses particle noise generated by debris adhering to the surface of recycled aggregate while enhancing the contrast between the old and new mortar interfaces.

[0039] In step 2, to address the recognition challenges posed by the highly irregular shapes of recycled aggregates, the multi-scale feature pyramid network model introduces dilated convolutions deep within the encoder. Dilated convolutions expand the receptive field without increasing the number of parameters by inserting gaps between convolution kernel elements. For large-sized recycled aggregates, the model can utilize dilated convolutions to capture their overall contour information, thereby forming more robust deep semantic features in the feature extraction stage of step 2. Furthermore, the residual correction module in the decoder is replaced with a recursive residual block, which optimizes the spatial information recovery process through multiple iterations, ensuring accurate fitting of irregular interfaces.

[0040] In step 3, the channel-spatial dual attention module adds a frequency domain analysis dimension. Before calculating the channel attention weights, the system performs a Fast Fourier Transform on the feature map to analyze the distribution of features in the frequency domain. Since cracks and pores in concrete often exhibit specific spatial frequency characteristics, the system enhances the sensitivity to capturing low-contrast targets by performing gain compensation on specific frequency bands in the frequency domain. The calculation of the spatial attention map introduces a multi-directional anisotropic convolution kernel to adapt to the geometric characteristics of the random distribution of cracks in recycled aggregate.

[0041] In step 4, the batch processing scheduling engine introduces an adaptive load balancing strategy based on task complexity. Because the image features of the recycled aggregate region are more complex, the computation time for its segmentation inference is often longer than that of the uniform region. The scheduling engine pre-evaluates the complexity of each batch of images (based on a textual description of grayscale entropy values) and assigns image blocks with higher complexity to computing cores with stronger computing power, ensuring synchronized parallel processing progress across the entire cluster and avoiding hardware idleness due to lagging individual tasks.

[0042] In step 5, the multi-dimensional quality assessment system adds a specific assessment index for the interface transition zone (ITZ) of recycled concrete. The system quantifies the segmentation accuracy of the interface zone by calculating the grayscale consistency within a specific pixel range extending outward from the aggregate edge and the smoothness of the segmentation mask. The topological integrity assessment is further refined into a connectivity analysis of the old mortar coating layer, ensuring that the old mortar and the original aggregate are correctly identified as a composite rather than separate individuals. In the structural fidelity index, the extraction of the centerline skeleton employs a multi-resolution pyramid strategy, calculating the skeleton offset at large, medium, and small scales and taking a weighted average to comprehensively reflect the morphological restoration effect of materials at different scales.

[0043] In step 6, the closed-loop optimization process incorporates a global search mechanism in the parameter space. When the quality assessment results of multiple consecutive batches fluctuate, the system not only fine-tunes the current model parameters but also initiates a lightweight Neural Architecture Search (NAS) subroutine to search for convolutional combinations more suitable for the current material properties within a pre-defined set of candidate operators. Through this deep closed-loop feedback, the system can continuously evolve as it accumulates processing tasks, achieving efficient and accurate analysis of complex recycled concrete data.

[0044] Example 3 This embodiment focuses on describing a 4D (three-dimensional spatial plus temporal dimension) CT image segmentation and evaluation method for the evolution process of concrete under in-situ loading. In this scenario, the concrete is subjected to continuously increasing pressure, and the internal cracks are in a state of dynamic expansion. The image quality may experience non-uniform degradation due to the occlusion of the loading device.

[0045] In the preprocessing stage of step 1, the system introduces a consistency constraint in the time dimension. For CT slices of the same specimen at different loading stages, the adaptive filtering algorithm not only refers to the standard deviation of the spatial neighborhood gradient of the current voxel, but also to the grayscale changes at the corresponding positions in the previous and next time steps. If the spatial domain shows a high gradient boundary, but the temporal domain shows that the abrupt grayscale change at that position does not conform to the crack propagation law, the system determines it to be random noise or an artifact and enhances the smoothing effect. This spatiotemporal joint filtering algorithm can significantly eliminate motion blur and artifact interference caused by dynamic loading.

[0046] In the model construction of step 2, the multi-scale feature pyramid network is extended into a recurrent convolutional network structure with memory capabilities. After extracting features at each level, the encoder stores the feature state in a Long Short-Term Memory (LSTM) convolutional module. This allows the network to "recall" the positions of aggregates and pores in the previous loading step when processing the image in the current loading step, thus maintaining extremely high structural continuity in the semantic segmentation results. The upsampling operation in the decoder part combines spatiotemporal interpolation techniques to ensure that while spatial resolution is restored, the deformation field in the temporal dimension is also smoothly processed.

[0047] In step 3, the attention guidance mechanism is configured as a triple attention mode, adding a temporal attention module to the original channel attention and spatial attention. The temporal attention module automatically identifies the crack initiation zone and expansion front caused by loading by comparing significant changes in the sequence images. The system assigns higher attention weights to regions with significant grayscale changes, enabling the model to refine the segmentation edges of these key dynamic regions, thereby achieving sub-pixel-level tracking of the microcrack evolution process.

[0048] In the batch scheduling of step 4, due to the exponential growth of 4D data volume, the scheduling engine employs hierarchical caching and pipeline prefetching techniques. The engine constructs the pipeline according to the temporal order of the loading stages. While the GPU is processing the image segmentation task in the k-th loading stage, the CPU has already completed the preprocessing of the k+1-th stage and prefetched the data into the system's cache. The heterogeneous computing architecture further supports distributed processing across nodes, distributing tasks to multiple computing nodes through a high-speed network. Each node is responsible for a different time segment, greatly compressing the total processing time of the entire data lifecycle.

[0049] In the quality assessment system of step 5, a time consistency score was added. This index is quantified by calculating the overlap rate and centroid offset between segmentation masks of adjacent time steps. If the spatial position of a component undergoes a non-physical jump within a very short time, the system will automatically lower the quality score of that batch. Topological integrity assessment focuses on the merging and bifurcation behavior of cracks, identifying the existence of erroneous topological connections by analyzing the monotonic change of the Euler number in the loading sequence. Structural fidelity introduces a deformation vector field (DVF) matching assessment, verifying the authenticity of the internal structure reconstruction by comparing the deformation predicted by the segmentation results with the surface deformation measured by digital image correlation (DIC) technology.

[0050] In the closed-loop optimization process of step 6, the system establishes a predictive adjustment model based on performance trends. The system collects quality assessment data from each loading step, uses regression analysis to predict potential image quality degradation trends during subsequent high-pressure stages, and adjusts filtering parameters and model weights in advance. For example, when it is predicted that the image signal-to-noise ratio will decrease by 20% as the load increases, the system automatically increases the baseline value of the adaptive filtering smoothing coefficient. This feedforward parameter adjustment mechanism, combined with a feedback-based fine-tuning procedure, ensures that the system can still produce stable and reliable data on the evolution of the internal structure of concrete under extreme experimental conditions.

[0051] In the above embodiments, all mathematical logic, numerical calculations, and parameter comparison processes involved have been converted into plain text descriptions. For example, all weighted summation operations performed through convolution kernels essentially involve locally weighting and integrating voxel regions using a specific spatial weight matrix; all threshold determination processes essentially involve dividing the target variable into intervals and selecting logical branches based on preset values. All logical connections, data flows, and closed-loop feedback in the system are completed in an automated process without human intervention, ensuring a high degree of objectivity and repeatability of the processing results.

[0052] In the implementation of the methods in the above embodiments, the internal data structure of the system adopts a highly optimized three-dimensional tensor storage format. Each three-dimensional voxel block is arranged in a contiguous address space in memory to maximize cache hit rate. For the local gradient calculation in step 1, the system utilizes the Single Instruction Multiple Data (SIMD) instruction set to achieve parallel acceleration on the CPU. In the model inference stages of steps 2 and 3, all weight matrices are stored in half-precision floating-point format, which reduces memory bandwidth pressure and improves computational efficiency while ensuring accuracy.

[0053] For concrete, a typical heterogeneous multiphase material, the system and method described in this invention solve the segmentation challenges caused by complex composition, overlapping gray levels, and massive amounts of data through adaptive preprocessing, multi-scale feature fusion, and closed-loop quality feedback. Specifically, by adjusting the smoothing coefficient in step 1, the convolutional layer depth in step 2, the attention weight in step 3, and the optimization threshold in step 6, this invention can flexibly adapt to the CT image analysis needs of various materials, from ordinary mortar to ultra-high performance concrete (UHPC), providing accurate geometric model support for subsequent material performance evaluation, finite element simulation, and durability analysis.

[0054] The specific implementation details of the quality threshold determination mechanism in the closed-loop optimization process described in this embodiment are as follows: The system pre-stores a feature library containing various typical concrete damage modes. When the quality score output in step 5 is lower than the set standard, the adjustment unit first quickly matches the current image features with the feature library to determine whether the main cause of the segmentation failure is excessive noise, insufficient contrast, or structural distortion. Based on the matching result, the system retrieves the corresponding optimization strategy from the parameter adjustment pool. If the main cause is insufficient contrast, the system will automatically enhance the histogram equalization intensity in step 1 and increase the contrast enhancement coefficient of the channel attention in step 3. If the main cause is structural distortion, the weight of the residual correction module in step 2 will be adjusted first. This targeted parameter adjustment scheme has higher convergence speed and stability compared to blind grid search or random fine-tuning.

[0055] Furthermore, the batch processing scheduling engine possesses robust exception handling and breakpoint resumption mechanisms during execution. When a hardware failure or memory overflow error is detected on a computing node, the engine immediately records the current progress status parameters and re-adds the failed task to the priority queue dispatcher, allowing other available nodes to take over. This robust design ensures that when processing ultra-large-scale CT image sequences lasting several days, the entire task will not crash due to local errors, greatly enhancing the system's engineering practicality.

[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

[0057] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for segmentation batch processing and quality assessment of concrete CT images, characterized in that, Includes the following steps: Step 1: Preprocess the original concrete CT image by using an adaptive filtering algorithm based on local gradient statistical characteristics to smooth the image voxels, suppress scattering noise and enhance the gray-level contrast between components, while preserving high-frequency detail information at the aggregate and pore boundaries. Step 2: Construct a multi-scale feature pyramid network model. Extract deep semantic features by downsampling layer by layer through the encoder, gradually restore spatial resolution by upsampling operation of the decoder, and fuse shallow geometric details of corresponding levels through skip connections to generate preliminary semantic segmentation results. Step 3: Introduce an attention guidance mechanism. After upsampling at each stage of the decoder, embed a channel-space dual attention module to dynamically adjust the feature response weights of different regions, enhance the attention to low-contrast targets, and optimize the segmentation boundary accuracy. Step 4: Establish a batch processing scheduling engine to automatically allocate computing resources based on the size, resolution, and grayscale distribution characteristics of the input image sequence, execute multiple segmentation tasks in parallel, and record intermediate state parameters during the processing. Step 5: Design a multi-dimensional quality assessment system to quantify the effectiveness of the segmentation results from three aspects: segmentation consistency, topological integrity, and structural fidelity. Step 6: Feed the quality assessment results back to the preprocessing and segmentation model parameter adjustment unit to form a closed-loop optimization process, dynamically adjusting the filtering intensity, attention weight distribution, and network learning rate based on the quality score of the current batch.

2. The method for segmentation, batch processing, and quality assessment of concrete CT images according to claim 1, characterized in that, The specific implementation process of the adaptive filtering algorithm used in step one is as follows: First, a three-dimensional sliding window centered on the current target voxel is defined. The size of the three-dimensional sliding window is set to a voxel cube of three times three times three or five times five times five, depending on the resolution of the image. Secondly, for all voxels within the three-dimensional sliding window, calculate the local average value of their grayscale values ​​and the standard deviation of their grayscale gradients. The calculation of the grayscale gradients covers the horizontal, vertical, and depth directions in the spatial coordinate system. Specifically, calculate the grayscale differences between the target voxel and its neighboring voxels in the three directions, and take the square root of the sum of the squares of the grayscale differences as the gradient magnitude. Then, a local smoothing coefficient is set based on the standard deviation of the gray-level gradient in the neighborhood of each voxel. When the calculated standard deviation of the gray-level gradient is lower than the preset global noise threshold, the region is determined to be inside the material, and a strong smoothing effect is applied to replace the gray value of the target voxel with the weighted average value in the neighborhood. Finally, when the standard deviation of the grayscale gradient is higher than the preset edge detection threshold, it is determined that there is a material boundary in the region, and the smoothing effect is weakened or the original grayscale value is kept unchanged to prevent the boundary blurring caused by the filtering operation.

3. The method for segmentation, batch processing, and quality assessment of concrete CT images according to claim 1, characterized in that, The construction and operation logic of the multi-scale feature pyramid network model in step two is as follows: The encoder is designed with five consecutive downsampling levels. Each level contains two sets of convolution operations with a kernel size of 3x3x3, and a stride convolutional layer with a stride of 2, which is used to reduce the spatial size of the feature map to half of its original size, while doubling the number of feature channels. The encoder part uses depthwise separable convolution to replace the traditional convolution operation. The depthwise separable convolution includes a channel-wise convolution stage and a pointwise convolution stage. In the channel-wise convolution stage, a convolution kernel is applied independently to each channel of the input feature map to extract spatial features. In the pointwise convolution stage, a one-to-one convolution kernel is used to linearly combine the outputs of all channels. The decoder performs a size recovery process symmetrical to that of the encoder, recovering the image size step by step through bilinear interpolation or transposed convolution operations. At each layer of the decoder, the feature maps from the corresponding layers of the encoder are directly stitched into the current feature stream through skip connections to trace the geometric contours in the shallow features.

4. The method for segmentation batch processing and quality assessment of concrete CT images according to claim 3, characterized in that, The multi-scale feature pyramid network model introduces a residual correction module at the end of the decoder. The residual correction module contains two residual branches. By learning the residual value between the prediction result and the real label, it compensates for the loss of spatial positioning information caused by multiple downsampling and upsampling operations. For complex concrete images involving recycled aggregates, the encoder introduces deep, dilated convolutions to expand the receptive field by inserting gaps between convolution kernel elements, capturing the overall contour information of large-sized recycled aggregates; the residual correction module in the decoder is replaced with a recursive residual block, which optimizes the spatial information recovery process through multiple iterations.

5. The method for segmentation, batch processing, and quality assessment of concrete CT images according to claim 1, characterized in that, The specific processing flow of the channel-space dual attention module in step three is as follows: First, global context information is aggregated along the channel dimension to generate channel attention maps. Specifically, global average pooling and global max pooling are performed on the input feature maps to obtain two vectors describing the global attributes of the channels. These vectors are then fed into a multilayer perceptron network with shared weights and processed by an activation function to obtain the channel attention weights. Secondly, the spatial attention map is calculated by combining the spatial saliency within the local window. The feature map after channel weight adjustment is subjected to average pooling and max pooling operations along the channel axis. The two two-dimensional spatial maps are then stitched together and a spatial attention mask is generated through a 7x7 convolutional layer and activation function. Finally, the channel attention weights are multiplied by the spatial attention mask, and the result of the multiplication is applied to the original feature map to achieve enhanced focus on key material regions and reduced response to background noise regions.

6. The method for segmentation batch processing and quality assessment of concrete CT images according to claim 5, characterized in that, The channel-space dual attention module also includes a frequency domain analysis dimension. Specifically, before calculating the channel attention weights, a fast Fourier transform is performed on the feature map to analyze the distribution pattern of features in the frequency domain, and gain compensation is performed on specific frequency bands in the frequency domain to enhance the sensitivity of capturing low-contrast targets. The calculation of the spatial attention map introduces a multi-directional anisotropic convolution kernel to adapt to the geometric characteristics of random distribution of cracks inside concrete. When dealing with dynamic images under loading conditions, the attention guidance mechanism is configured as a triple attention mode, adding a temporal attention module on the basis of channel attention and spatial attention, and locking the crack initiation zone and extension front caused by loading by comparing significant changes in the sequence images.

7. The method for segmentation batch processing and quality assessment of concrete CT images according to claim 1, characterized in that, The scheduling mechanism of the batch processing scheduling engine in step four is as follows: The batch processing scheduling engine supports heterogeneous computing architectures and uses a built-in resource monitoring program to monitor the load of the central processing unit, the video memory usage of the graphics processor, and the available memory space of the system in real time. The scheduling engine adopts a multi-threaded parallel mode, performing data prefetching, format decompression and preprocessing operations on the central processing unit, while deploying a deep learning inference model on the graphics processing unit. The scheduling engine dynamically adjusts the batch size of image slices sent to the graphics processor based on the current remaining amount of video memory. If sufficient video memory is detected, the batch size is increased. If the video memory is close to the warning line, the task is automatically split into smaller slice groups and multi-stream concurrency technology is started to achieve overlapping coverage of data transmission and computation inference. In addition, the scheduling engine introduces an adaptive load balancing strategy based on task complexity, which pre-evaluates the complexity of each batch of images based on grayscale entropy values ​​and prioritizes allocating image blocks with higher complexity to computing cores with stronger computing power.

8. The method for segmentation batch processing and quality assessment of concrete CT images according to claim 1, characterized in that, The quantitative logic of the multi-dimensional quality assessment system in step five is as follows: The segmentation consistency index includes the overall crossover ratio and the crossover ratio values ​​of the three sub-targets: aggregate, paste, and pore. The crossover ratio is calculated by the intersection volume between the voxel set predicted by the statistical model and the voxel set labeled by the benchmark, and the ratio of the intersection volume to the union volume of the two is calculated. The topological integrity assessment is performed by tracking the change in the number of connected components of each phase material before and after segmentation. The three-dimensional connected component labeling algorithm is used to count the number of each independent aggregate particle in the segmentation result and calculate the Euler number. When the change rate of the Euler number after segmentation exceeds the preset ratio, it is determined that there is over-segmentation or merging. For recycled concrete, a special evaluation index for the interface transition zone is added. The segmentation accuracy of the interface zone is quantified by calculating the grayscale consistency within a specific pixel range extending outward from the aggregate edge and the smoothness of the segmentation mask.

9. A method for segmentation, batch processing, and quality assessment of concrete CT images according to claim 8, characterized in that, The structural fidelity assessment is achieved by extracting the centerline skeleton from the edge gradient map of the original image and the edge map of the segmentation mask, respectively. The skeleton lines of the material are extracted using a thinning algorithm, and the average offset of the two sets of skeleton lines within a predetermined distance range in three-dimensional space is calculated. The smaller the average offset, the higher the restoration accuracy. For dynamic loading scenarios, a new time consistency scoring index has been added, which is quantified by calculating the overlap rate and centroid offset between segmentation masks of adjacent time steps; Structural fidelity also incorporates a matching degree evaluation of the deformation vector field, which verifies the authenticity of the internal structure reconstruction by comparing the deformation predicted by the segmentation results with the surface deformation measured by digital image correlation technology.

10. The method for segmentation batch processing and quality assessment of concrete CT images according to claim 1, characterized in that, The adjustment mechanism of the closed-loop optimization process in step six is ​​as follows: The closed-loop optimization process is equipped with a quality threshold determination mechanism. When the overall quality score of a certain batch is lower than the preset minimum qualified threshold, the model fine-tuning subroutine is triggered, which only updates the local layer parameters in the network that are related to the current quality problem. The adjustment unit pre-stores a feature library containing various typical concrete damage modes. When the quality score is lower than the standard, it matches the current image features with the feature library to determine whether the main cause of the segmentation failure is excessive noise, insufficient contrast or structural distortion. If the main cause is insufficient contrast, then increase the intensity of histogram equalization in preprocessing and increase the contrast enhancement coefficient of channel attention. If the primary cause is structural distortion, then the weights of the residual correction module should be adjusted first. For full lifecycle data, a predictive adjustment model based on performance trends is established. Evaluation data from each stage are collected for regression analysis to predict the image quality degradation trend in subsequent stages. Filtering parameters and model weights are adjusted in advance to form a feedforward adjustment mechanism.

Images