Steel slag concrete aggregate uniformity evaluation method and detection model construction method

Abstract

The invention relates to the technical field of building material evaluation and image detection, in particular to a steel slag concrete aggregate uniformity evaluation method and a detection model construction method. The method comprises the following steps: preparing to-be-detected concrete into a test piece, and curing; then cutting the test piece, and obtaining a cutting surface image as a concrete settlement image; dividing the settlement image into a plurality of regions, obtaining the steel slag area ratio and the natural aggregate area ratio in each region, and then substituting the steel slag area ratio and the natural aggregate area ratio into the following formula to calculate the aggregate uniformity coefficient K. According to the method, the substitution rate, the area ratio and the aggregate variation coefficient of the steel slag aggregate are measured and calculated through image segmentation, the aggregate uniformity of the steel slag concrete is evaluated, and concrete characterization is carried out through the K coefficient. Experiments prove that the variation trend of the K coefficient is related to the difference value of the area proportion indexes of the two aggregates, and the method is proved to be capable of accurately evaluating the uniformity of the aggregates in the steel slag concrete settlement image.

Description

Technical Field

[0001] The invention relates to the technical field of building material evaluation and image detection, in particular to a method for evaluating the uniformity of steel slag concrete aggregate and a method for constructing a detection model. Background Art

[0002] With the continuous growth of the world economy, the steel industry has developed rapidly. As a by-product of steelmaking, steel slag has a large output but a low utilization rate. Wasted steel slag not only causes waste of resources, but also causes environmental pollution. Replacing natural aggregate in concrete with steel slag can effectively improve the mechanical properties of concrete, while alleviating the shortage of natural resources and meeting the requirements of environmental protection. However, the density of steel slag aggregate is relatively high and it is easy to settle during the vibration process, thus affecting the uniformity of aggregate distribution. Poor aggregate uniformity will affect the strength and durability of concrete. Therefore, it is necessary to study the uniformity of steel slag coarse aggregate in concrete. Traditional methods for evaluating the uniformity of aggregate distribution mainly rely on visual judgment and manual measurement, which is inefficient and has large errors, affecting the objectivity and accuracy of the evaluation results. Therefore, it is necessary to develop an accurate, efficient and objective uniformity evaluation method.

[0003] Image analysis methods have been widely used to evaluate the characteristics of aggregates and concrete, significantly improving the accuracy and objectivity of evaluation and providing a more scientific means for concrete quality control. In concrete-related research, the use of segmentation technology in image analysis is an important means of evaluating aggregate uniformity. Although traditional image segmentation models such as threshold segmentation models, edge detection, and morphology provide certain solutions for aggregate recognition, they all rely on manually defined features and traditional machine learning models, and have limited accuracy when dealing with complex backgrounds and small targets.

[0004] Deep learning provides a new solution to the limitations of traditional segmentation models in concrete aggregate segmentation. The YOLO (You Only Look Once) series is widely used due to its simple structure and innovative full-image single regression technology, which not only achieves real-time performance but also outperforms traditional two-stage detectors in speed. The YOLOv8 model has been further optimized and improved based on the success of previous versions of the YOLO series. However, it is still difficult to accurately identify images with complex backgrounds and interference from many small targets.

[0005] 1(a) and 1(b), the YOLOv8 model includes a backbone network, a neck network, and a head network.

[0006] The backbone network includes a first convolutional network, a second convolutional network, a first C2F module, a third convolutional network, a second C2F module, a fourth convolutional network, a third C2F module, a fifth convolutional network, a fourth C2F module and an SPPF module connected in sequence;

[0007] The neck network includes sequentially connected: a first sampling layer, a first splicing module, a second sampling layer, a second splicing module, a first convolutional splicing module, and a second convolutional splicing module;

[0008] The head network includes three output networks, and the input ends of the three output networks are respectively connected to the output end of the second splicing module, the output end of the first convolution splicing module, and the output end of the second convolution splicing module. The output network identifies and labels steel slag and natural aggregate based on the output data corresponding to the neck network.

[0009] The output of the SPPF module is the output of the backbone network and is connected to the input of the neck network. The input of the first sampling layer and the input of the second convolutional splicing module are both connected to the input of the neck network.

[0010] The first splicing module and the second splicing module have the same structure and are collectively referred to as splicing modules. The splicing module is composed of a connected Concat splicing network and a C2f module. The input end of the Concat splicing network is the input end of the splicing module, the output end of the Concat splicing network is connected to the input end of the C2f module, and the output end of the C2f module is the output end of the splicing module. The input end of the first splicing module is also connected to the output end of the third C2F module, and the input end of the second splicing module is also connected to the output end of the second C2F module.

[0011] The first convolutional splicing module and the second convolutional splicing module have the same structure and are collectively referred to as convolutional splicing modules. The convolutional splicing module is composed of a connected convolutional network, a Concat splicing network and a C2f module. The input end of the convolutional network is the input end of the convolutional splicing module, the output end of the convolutional network is connected to the input end of the Concat splicing network, the output end of the Concat splicing network is connected to the input end of the C2f module, and the output end of the C2f module is the output end of the convolutional splicing module. The input end of the Concat splicing network in the first convolutional splicing module is also connected to the output end of the first splicing module; the input end of the Concat splicing network in the second convolutional splicing module is also connected to the output end of the SPPF module. Specifically, the convolutional network in the convolutional splicing module adopts a Conv network. Summary of the invention

[0012] In order to overcome the defect in the above-mentioned prior art that it is difficult to ensure the accuracy of aggregate recognition on settlement images due to the interference of complex image background and numerous small targets, the present invention proposes a steel slag concrete aggregate uniformity evaluation method, which can effectively reflect the uniformity changes of aggregates in steel slag concrete settlement images.

[0013] The present invention proposes a method for evaluating the uniformity of aggregates of steel slag concrete. First, the concrete to be tested is made into a test piece and cured. Then, the test piece is cut to obtain a cut surface image as a settlement image of the concrete. The settlement image is divided into multiple regions, and the steel slag area ratio and the natural aggregate area ratio in each region are obtained, and then substituted into the following formula to calculate the aggregate uniformity coefficient K.

[0014]

[0015] Where Ks represents the evaluation index of the slag area ratio on the settlement image, Kn represents the evaluation index of the natural aggregate area ratio on the settlement image; m represents the total number of regions on the settlement image; SAR i represents the slag area ratio of the ith region on the settlement image, NAR i represents the area ratio of natural aggregate in the ith region on the settlement image; Indicates the SAR of each area i The mean of Indicates the NAR of each region i The mean of; Ti represents the replacement rate of slag to natural aggregate in the ith region of the settlement image, Ti = SAR i / (SAR i +NAR i ); Ws represents the known slag replacement rate, Wn represents the natural aggregate retention rate, Wn = 1-Ws.

[0016] Preferably, the settlement image is a longitudinal cutting image of the specimen, and when calculating the slag area ratio and the natural aggregate area ratio in each area on the settlement image, the settlement image is divided into a plurality of equal-height areas from top to bottom.

[0017] Preferably, the settlement image is divided into multiple regions, and the method for obtaining the slag area ratio and the natural aggregate area ratio in each region is: aggregate recognition is performed on the settlement image through a target detection model to mark the slag and natural aggregate on the settlement image; and then the slag area ratio and the natural aggregate area ratio of each region set on the settlement image are calculated according to the marking results.

[0018] A detection model construction method proposed in the present invention is characterized in that it is used to construct a target detection model for annotating steel slag and natural aggregate on a settlement image; the detection model construction method first transforms the YOLOv8 model to obtain a basic model whose input is a settlement image and whose output is annotated results of steel slag and natural aggregate; then the basic model is made to perform machine learning on the constructed data set {settlement image; steel slag and natural aggregate annotated results}, and the converged basic model is used as the target detection model;

[0019] The method for transforming the YOLOv8 model includes: replacing the SPPF module in the trunk network with a WS module; the WS module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a pooling module, a fourth convolutional layer, a fifth convolutional layer, a first splicing network and a sixth convolutional layer connected in sequence, and also includes a seventh convolutional layer, the input end of the seventh convolutional layer is connected to the output end of the first convolutional layer, and the output end of the seventh convolutional layer is connected to the input end of the first splicing network; the first splicing network splices the output data of the seventh convolutional layer with the output data of the fifth convolutional layer and inputs them into the sixth convolutional layer; the output end of the sixth convolutional layer is the output end of the WS module, and the input end of the first convolutional layer is the input end of the WS module;

[0020] The pooling module includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a second splicing network; the input end of the first maximum pooling layer, the input end of the second maximum pooling layer and the input end of the third maximum pooling layer are all connected to the output end of the third convolutional layer, and the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer are all connected to the input end of the second splicing network; the second splicing network performs dimension sampling and splicing on the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer according to the set weights, and inputs the spliced ​​data to the fourth convolutional layer.

[0021] Preferably, the method for modifying the YOLOv8 model further comprises inserting an EMA module at the end of the trunk network.

[0022] A detection model construction method proposed in the present invention is used to construct a target detection model for steel slag and natural aggregate on annotated settlement images; the target detection model is obtained by modifying the YOLOv8 model, and the modification method is as follows: a trunk network of the YOLOv8 model is replaced with a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, an eighth convolution layer, a WS module and an EMA module that are sequentially connected; the output end of the EMA module is the output end of the trunk network;

[0023] The output end of the EMA module is respectively connected to the input end of the neck network and the input end of the third output network in the head network; the input end of the first splicing module in the neck network is connected to the output end of the fourth convolution module, and the output end of the fourth convolution module is also connected to the input end of the second output network of the head network; the input end of the second splicing module in the neck network is connected to the output end of the third convolution module, and the output end of the third convolution module is also connected to the input end of the first convolution splicing module and the input end of the first output network of the head network;

[0024] The WS module includes the first convolutional layer, the second convolutional layer, the third convolutional layer, the pooling module, the fourth convolutional layer, the fifth convolutional layer, the first splicing network and the sixth convolutional layer, which are connected in sequence, and also includes the seventh convolutional layer, the input end of the seventh convolutional layer is connected to the output end of the first convolutional layer, and the output end of the seventh convolutional layer is connected to the input end of the first splicing network; the first splicing network splices the output data of the seventh convolutional layer with the output data of the fifth convolutional layer and inputs them into the sixth convolutional layer; the output end of the sixth convolutional layer is the output end of the WS module, and the input end of the first convolutional layer is the input end of the WS module;

[0025] The pooling module includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a second splicing network; the input end of the first maximum pooling layer, the input end of the second maximum pooling layer and the input end of the third maximum pooling layer are all connected to the output end of the third convolutional layer, and the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer are all connected to the input end of the second splicing network; the second splicing network performs dimension sampling and splicing on the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer according to the set weights, and inputs the spliced ​​data to the fourth convolutional layer.

[0026] Preferably, the first convolution module, the second convolution module and the fourth convolution module have the same structure, and are composed of a sequentially connected convolution network, a normalized network and two ConNeXt V2 networks; the third convolution module is composed of a sequentially connected convolution network, a normalized network and six ConNeXt V2 networks.

[0027] A detection model building system proposed in the present invention includes a memory and a processor, wherein a computer program is stored in the memory, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the detection model building method.

[0028] The invention provides a steel slag concrete aggregate uniformity evaluation system, comprising a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the steel slag concrete aggregate uniformity evaluation method.

[0029] The present invention provides a storage medium storing a computer program and a target detection model. When the computer program is executed, it is used to implement the steel slag concrete aggregate uniformity evaluation method.

[0030] The advantages of the present invention are:

[0031] (1) The present invention proposes a method for evaluating the uniformity of aggregates in steel slag concrete. The method measures the replacement rate, area ratio and aggregate variation coefficient of steel slag aggregates by segmenting the image, thereby evaluating the uniformity of aggregates in steel slag concrete and characterizing it by the K coefficient. Experiments have shown that the K coefficient variation trend is correlated with the difference in the area ratio of the two aggregates, which proves that the method can accurately evaluate the uniformity of aggregates in the settlement image of steel slag concrete.

[0032] (2) The present invention identifies two types of aggregates on the settlement image based on a large model, and then calculates the area share. The YOLOv8 model is further improved for annotating settlement images. Experimental verification shows that the aggregate uniformity method proposed based on the improved model can objectively and accurately evaluate the uniformity of steel slag concrete aggregates, providing a new scientific method for concrete quality control. The present invention uses an instance segmentation model based on deep learning to achieve accurate segmentation of steel slag and natural aggregates, effectively solving the limitations of low precision of traditional segmentation models in complex environments and the problem that semantic segmentation cannot subdivide categories.

[0033] (3) The detection model construction method proposed in the present invention uses the WS module to replace the SPPF module, optimizes the spatial pyramid pool, and extracts more aggregate, pore and contour feature information by increasing the aggregate feature weight, thereby enhancing the local detail features and improving the segmentation accuracy of the aggregate. The WS module makes up for the lack of the attention mechanism, achieves a comprehensive improvement in all indicators, and verifies the effectiveness of the improved model.

[0034] (4) The WS module proposed in the present invention adopts a cascade method of pooling kernels to avoid the distortion and redundancy of feature extraction that may occur during image processing. By adding weights and integrating feature information processed by multiple standard convolutions, the model operation is accelerated without sacrificing the receptive field. Replacing the SPPF module in YOLOv8 with the WS module can effectively improve the model's ability to distinguish instance aggregates when processing small target recognition and segmentation tasks such as steel slag concrete aggregates, thereby improving recognition accuracy and real-time processing performance.

[0035] (5) An EMA module is added at the end of the backbone network to effectively perform fine pixel-level attention on high-level feature maps while keeping the channel dimension unchanged. This design reduces the depth of the model without significantly increasing the computational cost, thereby reducing the interference of non-critical features on the model and improving the generalization performance of the model in different application scenarios. The addition of the EMA attention mechanism enables the network to allocate more attention to important neurons, effectively avoiding the interference of non-essential features such as the background environment, thereby improving the accuracy of instance segmentation.

[0036] (6) In the present invention, by setting up the EMA module and embedding the attention mechanism into the WS module, the deep features can be better utilized, more aggregate feature information can be obtained, the key feature extraction capability can be improved, the computational efficiency can be optimized, the recognition accuracy and real-time processing capability can be improved, and the stability and generalization performance of the model can be enhanced.

[0037] (7) The improved model uses the Convnext V2-R network as the feature extraction network, thereby enhancing the feature extraction capability in the complex background of steel slag concrete. Comparative tests show that the proposed model is superior to similar segmentation models and can achieve objective and accurate instance segmentation in the case of complex background of steel slag concrete and many small targets. The segmentation effect is better than the existing aggregate image segmentation method in the environment with complex background and many small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1(a) is a module connection diagram of YOLOv8;

[0039] Figure 1(b) is the network topology diagram of YOLOv8;

[0040] Figure 2 This is a flow chart of a method for evaluating the uniformity of aggregates in steel slag concrete;

[0041] Figure 3 Flowchart of the method for building a target detection model;

[0042] Figure 4 This is the module connection diagram of YOLOv8-S;

[0043] Figure 5(a) is the network topology diagram of the WS module;

[0044] Figure 5(b) is a module connection diagram of the WS module;

[0045] Figure 6 This is the module connection diagram of YOLOv8-SE;

[0046] Figure 7(a) is a module connection diagram of YOLOv8-CSE;

[0047] Figure 7(b) is the network topology diagram of YOLOv8-CSE;

[0048] Figure 8 This is the module connection diagram of YOLOv8-C;

[0049] FIG9( a ) is a module connection diagram of the first convolution module;

[0050] FIG9( b ) is a module connection diagram of the third convolution module;

[0051] Fig.10 This is the module connection diagram of YOLOv8-CE;

[0052] Fig.11 This is the module connection diagram of YOLOv8-CW;

[0053] FIG12( a ) shows the variation trend of the accuracy of each model in the embodiment;

[0054] FIG12( b ) shows the variation trend of the recall rate of each model in the embodiment;

[0055] FIG12( c ) shows the variation trend of the average accuracy of each model in the embodiment;

[0056] FIG13( a ) shows the variation trend of the accuracy of YOLOv8-CSE and YOLOv8n in the embodiment;

[0057] FIG13( b ) shows the variation trend of recall rates of YOLOv8-CSE and YOLOv8n in the embodiment;

[0058] FIG13( c ) shows the variation trend of the average accuracy of YOLOv8-CSE and YOLOv8n in the embodiment;

[0059] Fig.14 The changing trend of slag area proportion with slag replacement rate;

[0060] Fig.15 is the changing trend of natural aggregate area proportion with slag replacement rate;

[0061] Fig.16 is the uniformity variation trend of Ks and Kn with different slag replacement rates Ws;

[0062] Fig.17 is the uniformity variation trend of coefficient K with different slag replacement rates Ws;

[0063] Figure 18(a) shows the sedimentation image when the slag replacement rate is 10%;

[0064] Figure 18(b) shows the sedimentation image when the slag replacement rate is 30%;

[0065] Figure 18(c) shows the sedimentation image when the slag replacement rate is 50%;

[0066] Figure 18(d) shows the sedimentation image when the slag replacement rate is 70%;

[0067] Figure 18(e) is a sedimentation image when the slag replacement rate is 90%. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Reference Figure 2 , a method for evaluating the uniformity of steel slag concrete aggregates proposed in this embodiment, firstly, the concrete to be tested is made into a test piece and cured;

[0070] Then the specimen is cut and the cut surface image is obtained as the settlement image of the concrete;

[0071] The settlement image is divided into multiple regions, the slag area ratio and natural aggregate area ratio in each region are obtained, and then substituted into the following formula to calculate the aggregate uniformity coefficient K.

[0072]

[0073] Where Ks represents the evaluation index of the slag area ratio on the settlement image, Kn represents the evaluation index of the natural aggregate area ratio on the settlement image; m represents the total number of regions on the settlement image; SAR i represents the slag area ratio of the ith region on the settlement image, NAR i represents the area ratio of natural aggregate in the ith region on the settlement image; Indicates the SAR of each area i The mean of Indicates the NAR of each region i The mean of; Ti represents the replacement rate of slag to natural aggregate in the ith region of the settlement image, Ti = SAR i / (SAR i +NAR i ); Ws represents the known slag replacement rate, Wn represents the natural aggregate retention rate, Ws = 1-Wn.

[0074] In this embodiment, aggregate recognition is performed on the settlement image through the target detection model to obtain slag and natural aggregate on the settlement image, and obtain the slag area ratio and natural aggregate area ratio of each area set on the settlement image.

[0075] The input of the target detection model is the settlement image, and the output is the detection results of steel slag and natural aggregate. Specifically, it can be expressed as a settlement image with steel slag and natural aggregate labeled, and then the steel slag area ratio and natural aggregate area ratio of each area set on the settlement image are calculated based on the labeled results. In specific implementation, a calculation layer can also be set at the end of the target detection model to calculate the steel slag area ratio SAR of each area set on the settlement image according to the labeled results output by the model. i , natural aggregate area ratio NAR i and slag replacement rate T i .

[0076] In this embodiment, for the target detection model including the computing layer, machine learning can be performed on the learning data set {settlement image; slag and natural aggregate labeling results}, and the converged basic model is used as the target detection model;

[0077] For the target detection model whose output is the segmentation result of steel slag and natural aggregate, that is, the target detection model that does not include a computational layer, machine learning can be performed on the learning data set {settlement image of concrete specimen with specified steel slag replacement rate; target segmentation result}. Specifically, the settlement image in the learning data set is annotated at the pixel level.

[0078] In specific implementation, the method for constructing the target detection model is as follows: first, a basic model and a learning data set are constructed, and the learning data set is used to store learning samples; then, a machine learning method is used to train the basic model on the learning data set, and the converged basic model is obtained as the target detection model.

[0079] In this implementation, the basic model is improved on the basis of the YOLOv8 model.

[0080] The first base model is denoted as YOLOv8-S.

[0081] In the baseline model YOLOv8, the first convolutional network, the second convolutional network, the convolutional network, the third convolutional network, the fourth convolutional network and the fifth convolutional network all use Conv networks; the convolutional network in the convolutional splicing module also uses Conv networks.

[0082] Reference Figure 4, Figure 5(a), Figure 5(b); Based on the YOLOv8 model, the YOLOv8-S model replaces the SPPF module in the trunk network with the WS module. The WS module includes the first convolutional layer, the second convolutional layer, the third convolutional layer, the pooling module, the fourth convolutional layer, the fifth convolutional layer, the first splicing network and the sixth convolutional layer connected in sequence. The WS module also includes the seventh convolutional layer, the input end of the seventh convolutional layer is connected to the output end of the first convolutional layer, and the output end of the seventh convolutional layer is connected to the input end of the first splicing network. The first splicing network splices the output data of the seventh convolutional layer with the output data of the fifth convolutional layer and inputs them into the sixth convolutional layer. The output end of the sixth convolutional layer is the output end of the WS module, and the input end of the first convolutional layer is the input end of the WS module.

[0083] The pooling module includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a second splicing network; the input end of the first maximum pooling layer, the input end of the second maximum pooling layer and the input end of the third maximum pooling layer are all connected to the output end of the third convolutional layer, and the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer are all connected to the input end of the second splicing network; the second splicing network performs 1 / 3 dimension sampling on the output data of the first maximum pooling layer, the output data of the second maximum pooling layer and the output data of the third maximum pooling layer, and splices them, and inputs the spliced ​​data into the fourth convolutional layer. Obviously, in the pooling module, the number of maximum pooling layers can also be set to other integers M greater than 1, and the second splicing network performs 1 / M dimension sampling on the output data of each maximum pooling layer and splices the output to ensure the stability of the data dimension and facilitate entry into the next layer of the network. Multiple maximum pooling layers are arranged in parallel, which is conducive to ensuring that data features are focused on from different angles, thereby ensuring a comprehensive analysis of the sedimentation image.

[0084] In a specific implementation, the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer all adopt a Conv network.

[0085] Specifically, the output data structures of the first maximum pooling layer, the second maximum pooling layer and the third maximum pooling layer in the WS module are the same, and the second splicing module extracts 1 / 3 of the output data dimensions of the first maximum pooling layer, the second maximum pooling layer and the third maximum pooling layer for splicing.

[0086] The second base model is denoted as YOLOv8-SE.

[0087] Reference Figure 6 ,Based on the YOLOv8-S model, the YOLOv8-SE model inserts an EMA module at the end of the trunk network, and the output of the WS module is connected to the input of the EMA module; the output of the EMA module is used as the output of the trunk network to connect to the input of the neck network.

[0088] The third basic model is denoted as YOLOv8-CSE.

[0089] Referring to Figure 7(a) and Figure 7(b), based on the YOLOv8-SE model, the trunk network of the model YOLOv8-CSE is modified. The trunk network of the YOLOv8-CSE model includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, an eighth convolution layer, a WS module and an EMA module connected in sequence; the output end of the EMA module is the output end of the trunk network. The output end of the EMA module is respectively connected to the input end of the first sampling layer in the neck network and the input end of the Concat splicing network in the second convolution splicing module, and the output end of the EMA module is also connected to the input end of the third output network in the head network.

[0090] The first convolution module, the second convolution module and the fourth convolution module have the same structure, and are composed of a sequentially connected convolution network, a normalized network and two ConNeXt V2 networks. The two ConNeXt V2 networks are sequentially connected to the output end of the normalized network, that is, the input end of the first ConNeXtV2 network is connected to the output end of the normalized network, and the input end of the second ConNeXt V2 network is connected to the output end of the first ConNeXtV2 network. In the first convolution module, the second convolution module and the fourth convolution module, the output end of the second ConNeXt V2 network is the output end of the convolution module.

[0091] The third convolutional module is composed of a convolutional network, a normalized network, and six ConNeXtV2 networks connected in sequence. The six ConNeXt V2 networks are sequentially connected to the output end of the normalized network, that is, the input end of the first ConNeXt V2 network is connected to the output end of the normalized network, the input end of the second ConNeXt V2 network is connected to the output end of the first ConNeXt V2 network, the input end of the third ConNeXtV2 network is connected to the output end of the second ConNeXtV2 network, the input end of the fourth ConNeXtV2 network is connected to the output end of the third ConNeXtV2 network, the input end of the fifth ConNeXtV2 network is connected to the output end of the fourth ConNeXt V2 network, and the input end of the sixth ConNeXtV2 network is connected to the output end of the fifth ConNeXt V2 network. In the sixth convolutional module, the output end of the sixth ConNeXtV2 network is the output end of the sixth convolutional module.

[0092] The convolutional networks in the first convolutional module, the second convolutional module, the third convolutional module and the fourth convolutional module all use Conv2d networks. The eighth convolutional layer also uses Conv2d networks.

[0093] The input end of the first splicing module in the neck network is connected to the output end of the fourth convolution module, and the output end of the fourth convolution module is also connected to the input end of the second output network of the head network. The input end of the second splicing module in the neck network is connected to the output end of the third convolution module, and the output end of the third convolution module is also connected to the input end of the first convolution splicing module and the input end of the first output network of the head network.

[0094] The fourth base model is denoted as YOLOv8-C.

[0095] Based on the YOLOv8 model, the trunk network of the model YOLOv8-C is modified. The trunk network of the YOLOv8-C model includes the first convolution module, the second convolution module, the third convolution module, the fourth convolution module, the eighth convolution layer and the SPPF module connected in sequence; the output of the SPPF module is the output of the trunk network. The output of the SPPF module is connected to the input of the first sampling layer in the neck network and the input of the Concat splicing network in the second convolution splicing module. The eighth convolution layer uses the Conv2d network.

[0096] The input end of the first splicing module in the neck network is connected to the output end of the fourth convolution module, and the output end of the fourth convolution module is also connected to the input end of the second output network of the head network. The input end of the second splicing module in the neck network is connected to the output end of the third convolution module, and the output end of the third convolution module is also connected to the input end of the first convolution splicing module and the input end of the first output network of the head network; the output end of the backbone network, i.e., the output end of the SPPF module, is also connected to the input end of the third output network in the head network.

[0097] In YOLOv8-C, the settings of the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module refer to the YOLOv8-CSE model.

[0098] The fifth basic model is denoted as YOLOv8-CE.

[0099] Based on the YOLOv8-C model, an EMA module is inserted at the end of the backbone network, and the output of the SPPF module is connected to the input of the EMA module; the output of the EMA module is used as the output of the backbone network, and the output of the EMA module is respectively connected to the input of the first sampling layer in the neck network and the input of the Concat splicing network in the second convolutional splicing module, and the output of the EMA module is also connected to the input of the third output network in the head network.

[0100] The sixth basic model is denoted as YOLOv8-CW.

[0101] Based on the YOLOv8-C model, the SPPF module in the backbone network is replaced by the WS module. The output of the WS module is the output of the backbone network. The output of the WS module is respectively connected to the input of the first sampling layer in the neck network, the input of the Concat network in the second convolutional concatenation module, and the input of the third output network in the head network.

[0102] The setting method of the WS module refers to the model YOLOv8-S.

[0103] The target detection model trained by the above-mentioned different basic models is verified in combination with specific embodiments below.

[0104] This embodiment builds an experimental system based on Windows 11 system and PyTorch2.0.0 (torch2.0.0, torchvison0.15.1) framework. The workstation is mainly configured with 13th Gen Intel(R) Core(TM) i7-13700HX5.00GHz CPU and NVIDIA GeForce RTX 4090 deep learning graphics GPU, Compute Unified DeviceArc (CUDA) 11.7. Based on Python3.9 platform, cuDNN is used to accelerate GPU training. All networks involved in training use the same environment configuration and hyperparameter settings. The specific training parameters are as follows: initial learning rate, batch size of the model, momentum parameter, weight decay, number of iterations, and image size are set to 0.01, 4, 0.937, 0.0005, 300epoch, and 640×640 pixels, respectively.

[0105] This embodiment constructs a steel slag concrete aggregate instance segmentation dataset. The dataset covers steel slag concrete aggregates with different substitution rates from 0% to 100%, with a gradient increase of 10% to form eleven groups, each containing 20 specimens. The test adopts the coarse aggregate volume replacement method to prepare steel slag concrete specimens, and the experimental mix ratio is determined according to the "Ordinary Concrete Mix Design Code" (JGJ 55-2011), see Table 1 for details.

[0106] Table 1 Steel slag concrete mix ratio (kg / m 3 )

[0107]

[0108] Each specimen was cut longitudinally after standard curing to obtain its settlement image. Anylabeling was used to interactively annotate the natural aggregate, slag aggregate and pores in the settlement image at the pixel level. Then, the annotated settlement image was cropped, rotated and enhanced by adding salt and pepper noise to obtain a total of 1784 images, which constituted the learning sample {settlement image; target segmentation result; slag area ratio SAR of each region i , natural aggregate area ratio NAR i and slag replacement rate T i}.

[0109] It is worth noting that in this embodiment, the slag area ratio SAR of each region in the learning sample i , natural aggregate area ratio NAR i and slag replacement rate T i is the value calculated based on the image area division and target segmentation results. The model in this embodiment is used to segment steel slag aggregate and natural aggregate from the settlement image. The higher the model accuracy, the closer the segmentation result is to the pixel-level labeled learning sample with its own label. i 、NAR i and T i It is calculated based on the target segmentation results of pixel-level annotation, which facilitates the subsequent verification of the effectiveness of the K coefficient.

[0110] The learning samples are divided into training set, validation set and test set in a ratio of 8:1:1 to ensure the comprehensiveness and balance of the data set. The training set, validation set and test set all contain learning samples under any slag substitution rate Ws. The comparison model and the basic model are used for machine learning on the training set and validation set. When the model iteration number reaches more than 300 times, the model is fixed, and then the model performance is tested on the test set.

[0111] The process of modeling machine learning on the training set includes the following steps:

[0112] St1, randomly select 200 training samples from the training set and substitute them into the model, allowing the model to learn the training samples;

[0113] St2. Randomly extract 25 verification samples from the verification set, let the model segment the settlement images in the verification samples, and calculate the model loss; this embodiment adopts CIOU loss, and SIOU or IOU loss can also be used;

[0114] St3, determine whether the model converges;

[0115] If not, back-propagate the iterative model parameters according to the model loss, and then return to step St1;

[0116] If yes, fix the model and complete the iteration.

[0117] In this embodiment, YOLOv5, YOLOv7 and multiple versions of YOLOv8 (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x) are used as comparison models.

[0118] YOLOv8n is a lightweight model suitable for devices with limited resources, such as embedded systems or mobile platforms. The model has a small amount of computation and a fast speed, but may sacrifice some detection accuracy. In this implementation, YOLOv8, which is used as the baseline model, is the YOLOv8n lightweight model.

[0119] YOLOv8s is slightly larger than YOLOv8sn, provides better performance, and is suitable for environments with medium computing resources;

[0120] YOLOv8m is a medium-sized model that balances speed and accuracy and is the first choice for many application scenarios;

[0121] YOLOv8l is a larger model that usually has significantly improved accuracy but requires more computing resources;

[0122] YOLOv8x is the largest model and generally achieves the best detection results, but also has the highest hardware requirements.

[0123] In this embodiment, a model YOLOv8-E is also constructed. Compared with the model YOLOv8, the EMA module is inserted at the end of the backbone network of YOLOv8-E, and the output end of the SPPF module is connected to the input end of the EMA module; the output end of the EMA module is connected to the input end of the first sampling layer in the neck network and the input end of the Concat splicing network in the second convolution splicing module as the output end of the backbone network.

[0124] In this implementation, both the target detection model and the comparison model are trained on the training set and the validation set, and then the model performance is tested on the test set.

[0125] In this embodiment, the test results of the comparison models YOLOv8n, YOLOv8-E and the target detection models YOLOv8-S, YOLOv8-CSE on the test set are shown in Table 2 below.

[0126] Table 2: Model test results (%)

[0127]

[0128] It is obvious that the performance of the model YOLOv8-E is improved in precision, recall and average accuracy compared with the baseline model YOLOv8n. This is because the addition of the EMA attention mechanism enables the network to allocate more attention to important neurons, effectively avoiding the interference of non-essential features such as the background environment, thereby improving the accuracy of instance segmentation. The three models YOLOv8-C, YOLOv8-CE and YOLOv8-CW proposed in this embodiment have greatly improved precision, recall and average accuracy compared with YOLOv8n and YOLOv8-E, proving that the present invention improves YOLOv8 and makes it more suitable for concrete aggregate detection.

[0129] The performance of the YOLOv8-S model is further improved compared to the YOLOv8-E model. It can be seen that the WS module can effectively improve the model's ability to distinguish instance aggregates when processing small target recognition and segmentation tasks such as steel slag concrete aggregates, thereby improving recognition accuracy and real-time processing performance.

[0130] Compared with the YOLOv8-S model, the YOLOv8-CSE model has improved the precision, recall and average accuracy by 5.8%, 4.8% and 3.5% respectively, and each performance has achieved a large and balanced improvement; compared with the baseline model YOLOv8n, the YOLOv8-CSE model has improved the precision, recall and average accuracy by 7.2%, 7.5% and 4.6% respectively. It can be seen that the segmentation effect of the YOLOv8-CSE model is better than the existing aggregate image segmentation method in the environment with complex backgrounds and many small targets, which verifies the effectiveness and significant advantages of the proposed model.

[0131] In this embodiment, during the training process, the indicator convergence curves of the model YOLOv8-CSE and each comparison model are shown in Figure 12(a), Figure 12(b), and Figure 12(c), respectively; It can be seen that the model YOLOv8-CSE shows obvious advantages over other models in various indicators. Specifically, the convergence speed of the model YOLOv8-CSE is significantly faster than that of other comparison models, and it can reach the convergence state in only 150 iterations. It is worth noting that even in the case of overfitting, its accuracy still exceeds that of other models. In addition, the recall rate and mAP50 performance are stable.

[0132] It can be seen that the model YOLOv8-CSE is superior to other models with instance segmentation capabilities in speed and accuracy in the steel slag concrete aggregate instance segmentation task, especially in small target detection and balancing samples of different categories and its adaptability to edge devices. Model YOLOv8-CSE achieves efficient instance segmentation with a smaller number of computational parameters, surpassing other large models. In the steel slag concrete aggregate instance segmentation, model YOLOv8-CSE shows a higher confidence score, reduces false detection and omission, proves its superiority over other models, and is more suitable for edge device deployment and engineering applications. Data analysis shows that compared with other models, model YOLOv8-CSE performs best in the three key evaluation indicators of precision, recall and mAP50, and its instance segmentation performance even exceeds larger models such as YOLOv7 and YOLOv8x.

[0133] The convergence curves of the indicators of the baseline models YOLOv8n and YOLOv8-CSE during the training process are shown in Figure 13(a), Figure 13(b), and Figure 13(c), respectively; it can be seen that the model YOLOv8-CSE surpassed YOLOv8n in the three detection indicators of precision, recall, and mAP50 at the beginning of training (about 5 iterations). Especially in the stable stage of training, the model YOLOv8-CSE showed a faster learning speed and better results, which further confirmed the significant advantages of the model YOLOv8-CSE in improving the accuracy of small target detection and solving the problem of sample imbalance, and effectively optimized the instance segmentation performance of the model.

[0134] In this embodiment, the aggregate uniformity coefficient K of steel slag concrete is calculated by combining the segmentation results of steel slag and natural aggregate by the model YOLOv8-CSE on the test set.

[0135] In this embodiment, the settlement image is divided into four vertical regions of equal height from top to bottom, and the pixel area distribution of aggregates at each level is calculated using OpenCV technology, thereby obtaining the slag area ratio SAR in each region. i Area ratio of natural aggregate NAR i Thus, the evaluation indexes Ks and Kn of the area ratio of steel slag to natural aggregate are calculated. In this embodiment, the area ratios of aggregates in four vertical regions under different steel slag replacement rates Ws are measured and calculated, and the results are shown in Table 3.

[0136] Table 3 Calculation results of area ratio of some steel slag concrete settlement images (%)

[0137]

[0138] In Table 2, number 10-1 represents learning sample 1 with Ws=10%, number 10-2 represents learning sample 2 with Ws=10%, number 30-1 represents learning sample 1 with Ws=30%, and so on. SAR 1 Represents the slag area ratio of the first area on the settlement image, SAR 1 Represents the area ratio of natural aggregate in the first area on the settlement image; SAR 2 Represents the slag area ratio of the second region on the settlement image, SAR 2 Represents the area ratio of natural aggregate in the second area on the settlement image; SAR 3 Represents the slag area ratio of the third region on the settlement image, SAR 3 Represents the area ratio of natural aggregate in the third region on the settlement image; SAR 4 Represents the slag area ratio of the fourth region on the settlement image, SAR 4 Represents the area ratio of natural aggregate in the fourth area on the settlement image.

[0139] Fig.14 and Fig.15 The distribution of steel slag aggregate and natural aggregate in various regions of the concrete settlement image is shown. Fig.16 The uniformity variation trend of Ks and Kn with different slag substitution rates Ws is shown; Fig.17 The uniformity variation trend of the comprehensive coefficient K is shown.

[0140] Example segmentation images under different slag replacement rates Ws are shown in Figures 18(a), 18(b), 18(c), 18(d), and 18(e). In the figures, green represents the distribution of slag and blue represents the distribution of natural aggregate.

[0141] from Fig.14 , Fig.15 From the above, it can be seen that when the replacement rate is 10%, the distribution of steel slag aggregate in each layer is relatively uniform, and there is no obvious aggregate sinking phenomenon. When the replacement rate is increased to 30%, the change in the area ratio of the aggregate shows that the steel slag aggregate gradually sinks, and the natural aggregate is concentrated in the upper and middle layers of the test block. The steel slag aggregate in samples 30-1 and 30-2 has a bottom sinking phenomenon, while the two aggregates in samples 30-3 and 30-4 are still evenly distributed in the middle and lower layers. When the replacement rate is 50%, the coarse aggregate is in the bottom layer (SAR 4 ) shows a denser distribution pattern, which is accompanied by aggregate segregation. When the replacement rate increases to 70%, the coarse aggregate mainly composed of steel slag forms a high-density layer at the bottom. As the replacement rate further increases to 90%, the distribution of steel slag in concrete becomes more stable and uniform. Fig.17 As shown, the actual situation of the sedimentation image is consistent with the changing trend of the area ratio calculation result.

[0142] from Figure 14-17 It can be seen that under different steel slag replacement rates, the change trend of K shows that the uniformity of steel slag concrete aggregate is mainly affected by the change in the uniformity of steel slag aggregate. When the steel slag replacement rate is 50%, the comprehensive uniformity of steel slag and natural aggregate is the best.

[0143] Of course, it is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any reference numerals in the claims should not be regarded as limiting the claims involved.

[0144] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

[0145] The techniques, shapes, and structural parts not described in detail in the present invention are all well-known techniques.

Claims

1. A method for evaluating the uniformity of steel slag concrete aggregate, characterized in that: First, the concrete to be tested is made into a specimen and cured; then the specimen is cut to obtain the cut surface image as the settlement image of the concrete; the settlement image is divided into multiple regions, the slag area ratio and the natural aggregate area ratio in each region are obtained, and then substituted into the following formula to calculate the aggregate uniformity coefficient K; Where Ks represents the evaluation index of the slag area ratio on the settlement image, Kn represents the evaluation index of the natural aggregate area ratio on the settlement image; m represents the total number of regions on the settlement image; SAR i represents the slag area ratio of the ith region on the settlement image, NAR i represents the area ratio of natural aggregate in the ith region on the settlement image; Indicates the SAR of each area i The mean of Indicates the NAR of each region i The mean of; Ti represents the replacement rate of slag to natural aggregate in the ith region of the settlement image, Ti = SAR i / (SAR i +NAR i ); Ws represents the known slag replacement rate, Wn represents the natural aggregate retention rate, Wn = 1-Ws.

2. The method for evaluating uniformity of steel slag concrete aggregate according to claim 1, characterized in that: The settlement image is a longitudinal cutting image of the specimen. When calculating the slag area ratio and the natural aggregate area ratio in each area on the settlement image, the settlement image is divided into multiple equal-height areas from top to bottom.

3. The method for evaluating uniformity of steel slag concrete aggregate according to claim 1, characterized in that: The settlement image is divided into multiple regions, and the method for obtaining the slag area ratio and the natural aggregate area ratio in each region is as follows: aggregate recognition is performed on the settlement image through the target detection model to mark the slag and natural aggregate on the settlement image; and then the slag area ratio and the natural aggregate area ratio of each region set on the settlement image are calculated according to the marking results.

4. A method for constructing a detection model for the method for evaluating the uniformity of steel slag concrete aggregate according to claim 3, characterized in that: Used to construct a target detection model for annotating steel slag and natural aggregate on a settlement image; the detection model construction method first transforms the YOLOv8 model to obtain a basic model whose input is a settlement image and whose output is the annotation results of steel slag and natural aggregate; then let the basic model be used in the constructed data set {settlement image; Machine learning is performed on the steel slag and natural aggregate labeling results, and the converged basic model is used as the target detection model; The method of modifying the YOLOv8 model includes: replacing the SPPF module in the trunk network with the WS module; The WS module includes the first convolutional layer, the second convolutional layer, the third convolutional layer, the pooling module, the fourth convolutional layer, the fifth convolutional layer, the first splicing network and the sixth convolutional layer, which are connected in sequence, and also includes the seventh convolutional layer, the input end of the seventh convolutional layer is connected to the output end of the first convolutional layer, and the output end of the seventh convolutional layer is connected to the input end of the first splicing network; the first splicing network splices the output data of the seventh convolutional layer with the output data of the fifth convolutional layer and inputs them into the sixth convolutional layer; the output end of the sixth convolutional layer is the output end of the WS module, and the input end of the first convolutional layer is the input end of the WS module; The pooling module includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a second splicing network; the input end of the first maximum pooling layer, the input end of the second maximum pooling layer and the input end of the third maximum pooling layer are all connected to the output end of the third convolutional layer, and the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer are all connected to the input end of the second splicing network; the second splicing network performs dimension sampling and splicing on the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer according to the set weights, and inputs the spliced ​​data to the fourth convolutional layer.

5. The detection model construction method according to claim 4, characterized in that: The method of modifying the YOLOv8 model also includes inserting an EMA module at the end of the trunk network.

6. A method for constructing a detection model for the method for evaluating the uniformity of steel slag concrete aggregate according to claim 3, characterized in that: A target detection model for annotating steel slag and natural aggregate on the settlement image is constructed; the target detection model is modified on the YOLOv8 model, and the modification method is as follows: the trunk network of the YOLOv8 model is replaced with the first convolution module, the second convolution module, the third convolution module, the fourth convolution module, the eighth convolution layer, the WS module and the EMA module connected in sequence; the output end of the EMA module is the output end of the trunk network; The output end of the EMA module is respectively connected to the input end of the neck network and the input end of the third output network in the head network; the input end of the first splicing module in the neck network is connected to the output end of the fourth convolution module, and the output end of the fourth convolution module is also connected to the input end of the second output network of the head network; the input end of the second splicing module in the neck network is connected to the output end of the third convolution module, and the output end of the third convolution module is also connected to the input end of the first convolution splicing module and the input end of the first output network of the head network; The WS module includes the first convolutional layer, the second convolutional layer, the third convolutional layer, the pooling module, the fourth convolutional layer, the fifth convolutional layer, the first splicing network and the sixth convolutional layer, which are connected in sequence, and also includes the seventh convolutional layer, the input end of the seventh convolutional layer is connected to the output end of the first convolutional layer, and the output end of the seventh convolutional layer is connected to the input end of the first splicing network; the first splicing network splices the output data of the seventh convolutional layer with the output data of the fifth convolutional layer and inputs them into the sixth convolutional layer; the output end of the sixth convolutional layer is the output end of the WS module, and the input end of the first convolutional layer is the input end of the WS module; The pooling module includes: a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer and a second splicing network; the input end of the first maximum pooling layer, the input end of the second maximum pooling layer and the input end of the third maximum pooling layer are all connected to the output end of the third convolutional layer, and the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer are all connected to the input end of the second splicing network; the second splicing network performs dimension sampling and splicing on the output end of the first maximum pooling layer, the output end of the second maximum pooling layer and the output end of the third maximum pooling layer according to the set weights, and inputs the spliced ​​data to the fourth convolutional layer.

7. The detection model construction method according to claim 6, characterized in that: The first convolution module, the second convolution module and the fourth convolution module have the same structure, and are composed of a sequentially connected convolution network, a normalized network and two ConNeXtV2 networks; the third convolution module is composed of a sequentially connected convolution network, a normalized network and six ConNeXtV2 networks.

8. A detection model building system, characterized in that: It includes a memory and a processor, the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the detection model construction method as described in any one of claims 4 to 7.

9. A steel slag concrete aggregate uniformity evaluation system, characterized in that: It comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the steel slag concrete aggregate uniformity evaluation method as described in any one of claims 1, 2 or 3.

10. A storage medium, characterized in that: A computer program and a target detection model are stored, and when the computer program is executed, it is used to implement the steel slag concrete aggregate uniformity evaluation method as described in any one of claims 1, 2 or 3.

Images