CN118036476A - Precast concrete crack detection model, method, system and readable medium - Google Patents

Abstract

The invention relates to the technical field of precast concrete surface detection and computers, in particular to a precast concrete crack detection model, a precast concrete crack detection method, a precast concrete crack detection system and a precast concrete crack detection readable medium. According to the training method of the precast concrete crack detection model, a YOLOv model is optimized firstly, and then machine learning is carried out. According to the invention, an LW module is set for three paths of data splicing, wherein one path is data input by the LW module, the LW module participates in the splicing, a larger characteristic diagram is reserved, and the detection of a small target is facilitated; the other path is data which is subjected to average pooling and maximum pooling and then dimension superposition on the data input by the LW module, the average pooling is beneficial to extracting overall characteristics, the maximum pooling is beneficial to extracting the most prominent characteristics, and the combination of the two ensures the overall characteristic extraction of the data, thereby being beneficial to improving the detection precision.

Description

Precast concrete crack detection model, method, system and readable medium

Technical Field

The invention relates to the fields of precast concrete surface detection and computer technology, in particular to a precast concrete crack detection model, method, system and readable medium.

Background technique

Quality control is becoming more and more important in product production. Surface defects of products affect their quality. A large part of the cost of building construction is spent on rework caused by material defects. Crack defects are a very common type of surface defects of prefabricated components. Accurate crack defect detection plays a very important role in improving product quality and reducing construction costs.

At present, the main methods for detecting and evaluating the quality of prefabricated components include manual recognition detection methods, detection methods based on image processing algorithms, and deep learning methods. Among them, the traditional manual detection method has a low sampling rate, the detection accuracy is greatly affected by subjective factors such as manual experience and fatigue, the labor intensity is high, the detection efficiency is low, and the real-time performance is poor. The image processing algorithm detection method mainly detects cracks in images of the same material and texture background, and it is currently impossible to directly detect cracks in color images. The existing deep learning algorithm detection method generally compresses the image to be detected directly to a smaller size to meet the neural network input requirements, and then sends it to the pre-trained neural network model for detection to obtain the prediction result. When processing high-resolution crack images, this type of algorithm will cause image distortion due to excessive compression and have a high false detection rate.

Summary of the invention

In order to overcome the defects of low efficiency or low precision in the above-mentioned prior art of crack detection of precast concrete parts, the present invention proposes a training method for a precast concrete crack detection model, which can train an efficient and accurate precast concrete crack detection model.

The present invention proposes a method for training a precast concrete crack detection model, comprising the following steps:

First, a basic model and learning samples are obtained. The learning samples are the associated crack shooting images and crack annotated images. The basic model is based on the YOLOv5 model, and the connected Upsample unit and Concat unit are replaced by the LW module. The LW module first performs pooling on the input data of the Upsample unit in the YOLOv5 model, and then performs three-way splicing of the pooled data with the two inputs of the Upsample unit and Concat unit combination module in the YOLOv5 model and outputs them.

Then the basic model is made to learn the learning samples to iterate the model parameters until convergence; the converged basic model is used as a precast concrete crack detection model, whose input is the captured image and output is the crack annotation image.

Preferably, the LW module includes an average pooling layer, a maximum pooling layer, a dimension superposition unit and a fifth Concat unit; the fifth Concat unit is provided with three input terminals; the LW module is provided with two input terminals; in the Upsample unit and the Concat unit combination module in the YOLOv5 model, the input terminal of the Upsample unit is recorded as the first input terminal of the combination module, and the input terminal of the Concat unit connected to the second C3 module is recorded as the second input terminal of the combination module; the first input terminal of the LW module replaces the first input terminal of the combination module, and the second input terminal of the LW module replaces the second input terminal of the combination module;

The first input end of the LW module is respectively connected to the input end of the average pooling layer, the input end of the maximum pooling layer and the first input end of the fifth Concat unit, the output end of the average pooling layer and the output end of the maximum pooling layer are both connected to the input end of the dimension superposition unit, the output end of the dimension superposition unit is connected to the second input end of the fifth Concat unit, and the third input end of the fifth Concat unit is connected to the second input end of the LW module; the output end of the fifth Concat unit serves as the output end of the LW module.

Preferably, in the YOLOv5 model, modules with the same structure are named sequentially along the direction of data flow; based on the YOLOv5 model, the basic model also replaces the first C3 module and the fourth C3 module in the data flow direction with an MMT module, and the MMT module includes a Bottleneck unit, a MultiAttenCat unit and a tenth Conv unit connected in sequence; the input end of the MMT module is respectively connected to the input end of the Bottleneck unit and the input end of the MultiAttenCat unit, and the output end of the tenth Conv unit is used as the output end of the MMT module.

Preferably, the basic model is based on the YOLOv5 model, and the fourth Conv unit and the fifth Conv unit are both replaced by RepVGG units.

Preferably, the iteration process of the model parameters comprises the following steps:

St1, divide the learning samples into training set and test set;

St2, extract multiple samples from the training set as training samples, and let the basic model learn the training samples to iterate the model parameters;

St3, extract multiple samples from the test set as test samples, let the basic model predict the test samples and output crack annotation images;

St4, calculate the loss of the basic model on the test sample to determine whether the basic model converges; if not, return to step St2; if yes, fix the basic model.

Preferably, in St4, the loss of the basic model is cross entropy loss or mean square error loss.

A precast concrete crack detection method proposed in the present invention is characterized in that the detection model is firstly obtained by adopting the training method of the precast concrete crack detection model; then the target to be detected is photographed to obtain a target image; then the target image is input into the detection model, and the detection model outputs a crack annotation image to annotate the crack type.

Preferably, the crack types include: transverse cracks, longitudinal cracks and fatigue cracks.

A precast concrete crack detection 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 training method of the precast concrete crack detection model.

The present invention provides a readable medium storing a computer program, which is used to implement the training method of the precast concrete crack detection model when executed. The advantages of the present invention are:

(1) In the present invention, an LW module is set to perform three-way data splicing, one of which is the data input by the LW module. It participates in the splicing and retains a larger feature map, which is beneficial to the detection of small targets; the other is the data input by the LW module after average pooling and maximum pooling and dimension superposition. Average pooling is beneficial to extracting overall features, and maximum pooling is beneficial to extracting the most prominent features. The combination of the two ensures the comprehensive feature extraction of the data, which is beneficial to improving the detection accuracy.

(2) The MMT module improves data processing speed by replacing the C3 module in the original YOLOv5 with a lightweight network model. The RepVGG module is used to reduce the network weight by reusing a simple convolutional structure, and the calculation of the convolutional layer is equivalently expressed as the weighted sum of several small convolutional blocks. The MMT module is used to optimize the network structure, the RepVGG module is used to reduce the network weight, and the LW module improves the detection accuracy.

(3) The detection model proposed in the present invention can detect different types of cracks on precast concrete parts. Using the detection model proposed in the present invention to perform surface detection of precast concrete parts can greatly improve the detection accuracy of various types of cracks and improve the detection efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a first detection model construction method;

Figure 2 is a diagram of the traditional YOLOv5 network structure;

FIG3 is a network structure diagram of the first detection model;

Figure 4 is a diagram of the LW module network structure;

FIG5 is a diagram of the MMT module network structure;

FIG6 is a diagram of the network structure of the second detection model;

FIG7 is a network structure diagram of the third detection model;

Figure 8 shows images of three concrete cracking modes;

Figure 9 is the loss curve of the first detection model;

Figure 10 shows the loss curve of the second detection model;

Figure 11 shows the loss curve of the third detection model;

Figure 12 is a comparison of P-R curves of various models;

Figure 13 is the accuracy confusion matrix of the first detection model;

Figure 14 is the accuracy confusion matrix of the second detection model;

Figure 15 is the accuracy confusion matrix of the third detection model.

Detailed ways

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.

The traditional YOLOv5 model includes: backbone network (backbone), neck network (neck) and head network (head);

2 , the backbone network includes a first Conv unit, a second Conv unit, a first C3 module, a third Conv unit, a second C3 module, a fourth Conv unit, a third C3 module, a fifth Conv unit, a fourth C3 module and an SPPF module connected sequentially; the input end of the first Conv unit serves as the input end of the backbone network and is also the input end of the YOLOv5 model;

The neck network includes a sixth Conv unit, a first Upsample unit, a first Concat unit, a fifth C3 module, a seventh Conv unit, a second Upsample unit, a second Concat unit, a sixth C3 module, an eighth Conv unit, a third Concat unit, a seventh C3 module, a ninth Conv unit, a fourth Concat unit, and an eighth C3 module connected sequentially;

The output end of the SPPF module is connected to the input end of the sixth Conv unit, the input end of the first Concat unit is also connected to the output end of the third C3 module, the input end of the second Concat unit is also connected to the output end of the second C3 module, the input end of the third Concat unit is also connected to the output end of the seventh Conv unit, and the input end of the fourth Concat unit is also connected to the output end of the sixth Conv unit;

The head network includes: a first Detect unit, a second Detect unit and a third Detect unit; the input end of the first Detect unit is connected to the output end of the sixth C3 module, the input end of the second Detect unit is connected to the output end of the seventh C3 module, and the input end of the third Detect unit is connected to the output end of the eighth C3 module;

The first Detect unit outputs a first detection result, the second Detect unit outputs a second detection result, and the third Detect unit outputs a third detection result.

In actual application, one or more of the first detection result, the second detection result and the third detection result can be selected as the output of the YOLOv5 model as needed.

The data form H×P×Q is defined to represent the H-dimensional data representation of the feature graph of size P×Q. In this embodiment, let H= ^2h+1 , 2h ⁺² , 2h ⁺³ , 2h ⁺⁴ , n, 2n, i, 4i; P=x, x/2, x/4, x/8, x/16, p, p/2, 2p, j; Q=y, y/2, y/4, y/8, y/16, q, q/2, 2q, k; h, x, y, n, p, q, i, j and k are all set values.

In the YOLOv5 model, the first Conv unit performs convolution processing on the input image data and outputs data ^2h ×x×y; the data ^2h ×x×y is converted into data 2h ⁺¹ ×(x/2)×(y/2) by the second Conv unit convolution processing, and the data 2h ⁺¹ ×(x/2)×(y/2) is processed by the first C3 module and the third Conv unit in turn, and the third Conv unit outputs data 2h ⁺² ×(x/4)×(y/4); the data 2h ⁺² ×(x/4)×(y/4) is processed by the second C3 module and the fourth Conv unit in turn, and the fourth Conv unit outputs data 2h ⁺³ ×(x/8)×(y/8); the data 2h ⁺³ ×(x/8)×(y/8) is processed by the third C3 module and the fifth Conv unit in turn, and the fifth Conv unit outputs data 2h ⁺⁴ ×(x/16)×(y/16) ^; ×(x/16)×(y/16) is processed by the fourth C3 module, the SPPF module and the sixth Conv unit in sequence and then outputs the data 2h ⁺³ ×(x/16)×(y/16).

The data ^2h+3 ×(x/16)×(y/16) output by the sixth Conv unit is converted into data 2h ⁺³ ×(x/8)×(y/8) by the first Upsample unit; the first Concat unit concatenates the data 2h ⁺³ ×(x/8)×(y/8) output by the third C3 module and the data 2h ⁺³ ×(x/8)×(y/8) output by the first Upsample unit.

The data 2h ⁺⁴ ×(x/8)×(y/8) output by the first Concat unit is converted into data 2h ⁺³ ×(x/8)×(y/8) by the fifth C3 module and then input into the seventh Conv unit; the seventh Conv unit converts the input data into data 2h ⁺² ×(x/8)×(y/8) and then inputs into the second Upsample unit.

The second Concat unit concatenates the data 2h ⁺² ×(x/4)×(y/4) output by the second C3 module and the data 2h ⁺² ×(x/4)×(y/4) output by the second Upsample unit. The data 2h ⁺³ ×(x/4)×(y/4) output by the second Concat unit is converted into data 2h ⁺² ×(x/4)×(y/4) by the sixth C3 module and then input into the eighth Conv unit. The eighth Conv unit converts the input data into data 2h ⁺² ×(x/8)×(y/8) and then inputs into the third Concat unit.

The third Concat unit concatenates the data 2h ⁺² ×(x/8)×(y/8) output by the seventh Conv unit and the data 2h ⁺² ×(x/8)×(y/8) output by the eighth Conv unit into data 2h ⁺³ ×(x/8)×(y/8). The data 2h ⁺³ ×(x/8)×(y/8) output by the third Concat unit is processed by the seventh C3 module and the ninth Conv unit in sequence, and the ninth Conv unit outputs data 2h ⁺³ ×(x/16)×(y/16).

The fourth Concat unit concatenates the data 2h ⁺³ ×(x/16)×(y/6) output by the sixth Conv unit and the data 2h ⁺³ ×(x/16)×(y/16) output by the ninth Conv unit into data 2h ⁺⁴ ×(x/16)×(y/16), and the eighth C3 module processes the data 2h ⁺⁴ ×(x/16)×(y/16) output by the fourth Concat unit.

The precast concrete crack detection model proposed in the present invention, referred to as the detection model, is used to detect cracks on the surface of precast concrete; the input of the detection model is: a photographed image of the precast concrete, and the output is: a crack annotation image, that is, an image with annotated cracks.

The first detection model

1 , 3 and 4 , the first detection model proposed in the present invention is obtained by improving the traditional YOLOv5 model, and the improvement method includes the following steps S1-S2.

S1. Construct an LW module, which includes an average pooling layer (mean-pooling), a maximum pooling layer (max-pooling), a dimension superposition unit and a fifth Concat unit; the fifth Concat unit is provided with three input terminals; the LW module is provided with two input terminals, the first input terminal of the LW module is respectively connected to the input terminal of the average pooling layer, the input terminal of the maximum pooling layer and the first input terminal of the fifth Concat unit, the output terminal of the average pooling layer and the output terminal of the maximum pooling layer are both connected to the input terminal of the dimension superposition unit, the output terminal of the dimension superposition unit is connected to the second input terminal of the fifth Concat unit, and the third input terminal of the fifth Concat unit is connected to the second input terminal of the LW module; the output terminal of the fifth Concat unit serves as the output terminal of the LW module.

In this way, the data n×p×q input from the first input end of the LW module are processed by average pooling and maximum pooling respectively, and then dimensionally superimposed. The dimensionally superimposed data n×(p/2)×(q/2) and the data input from the two input ends of the LW module are dimensionally spliced through the fifth Concat unit as the output data 2n×2p×2q of the LW module. In the LW module, the fifth Concat unit performs weighted splicing on the 3 input data, and the parameter weights W ₁ , W ₂ and W ₃ on the three channels are all set values.

S2. Replace the first Upsample unit and the first Concat unit in the YOLOv5 model with an LW module, recorded as the first LW module; replace the second Upsample unit and the second Concat unit in the YOLOv5 model with an LW module, recorded as the second LW module; the first input end of the first LW module is connected to the output end of the sixth Conv unit, the second input end of the first LW module is connected to the output end of the third C3 module, and the output end of the first LW module is connected to the input end of the fifth C3 module; the first input end of the second LW module is connected to the output end of the seventh Conv unit, the second input end of the second LW module is connected to the output end of the second C3 module, and the output end of the second LW module is connected to the input end of the sixth C3 module.

In this embodiment, the LW module performs three-way data splicing, one of which is the data input to the LW module. It participates in the splicing and retains a larger feature map, which is beneficial to the detection of small targets. The other is the data input to the LW module after average pooling and maximum pooling and dimension superposition. Average pooling is beneficial to extracting overall features, and maximum pooling is beneficial to extracting the most prominent features. The combination of the two ensures comprehensive feature extraction of the data, which is beneficial to improving detection accuracy.

The second detection model

5 and 6 , the second detection model proposed in the present invention is obtained by further improving the first detection model in that the first C3 module and the fourth C3 module are replaced by the first MMT module and the second MMT module respectively.

The first MMT module and the second MMT module have the same structure and are collectively referred to as MMT modules, which include a Bottleneck unit, a MultiAttenCat unit and a tenth Conv unit connected in sequence; the input end of the MMT module is respectively connected to the input end of the Bottleneck unit and the input end of the MultiAttenCat unit, and the output end of the tenth Conv unit serves as the output end of the MMT module.

The input data i×j×k of the MMT module is converted into data 4i×j×k by the Bottleneck unit; the MultiAttenCat unit samples and splices the input data i×j×k and the data 4i×j×k to generate data i×j×k as output data.

This embodiment introduces a lighter MMT module in the backbone network to replace two C3 modules to increase data processing speed.

The main function of the C3 module is to improve the information flow in the model through partial connections across stages; the C3 module convolves the input data and then inputs it into the Bottleneck network for processing, and the output of the Bottleneck network is concatenated with the input data.

The MMT module in this embodiment splices the processing result of the Bottleneck network on the input data with the input data through the MultiAttenCat unit to integrate more information. The MultiAttenCat unit can manually set the weights of different inputs so that the network can most adaptably select channels and weights, and can suppress features that interfere with the output, thereby facilitating the detection of targets.

The third detection model

7 , the third detection model proposed in the present invention is obtained by further improving the second detection model in that the fourth Conv unit and the fifth Conv unit are both replaced by RepVGG units.

The fourth Conv unit and the fifth Conv unit are the two convolution modules with the largest number of parameters in the original YOLOv5 model. In this embodiment, the fourth Conv unit and the fifth Conv unit are lightweighted by replacing the RepVGG unit, thereby improving the accuracy of the model while reducing the number of parameters.

The three detection models mentioned above are verified below in combination with specific embodiments.

In this embodiment, the data set for target detection includes 1371 concrete damage images taken with a high-resolution camera. These concrete damage images are taken under different lighting conditions and depict concrete pavements with different flatness. In this embodiment, Photoshop is used to manually annotate each crack image as a binary image and crop it, and a resolution of 512×512 is obtained after cropping. That is, the samples in this embodiment are: concrete damage images with crack images annotated, and the crack images are binary images with a resolution of 512×512.

Concrete cracking includes three modes, namely longitudinal cracking, transverse cracking and fatigue cracking, also recorded as longitudinal cracks, transverse cracks and fatigue cracks, as shown in Figure 8.

In this embodiment, the data set is randomly divided into a training set and a validation set. The training set and the validation set contain 960 and 411 different crack images, respectively, as shown in Table 1.

Table 1 The number of images of each damage type in the self-built dataset

In this embodiment, the model evaluation metrics include: accuracy, recall, mean average precision (mAP), F1 score, GFLOPs, and number of parameters (Params).

The larger the precision, recall, and mAP values, the more accurate the crack detection is, and the smaller the Params and GFLOPs values, the less computing power the model requires. For F1, a value of 1 indicates the best performance.

In this embodiment, the first detection model is recorded as YOLOv5-L, the second detection model is recorded as YOLOv5-ML, and the third detection model is recorded as YOLOv5-RML. In this embodiment, the models Fast R-CNN, YOLOv5, YOLOv6, YOLOv7 and YOLOv8 are also selected as comparison models, so as to verify the performance of the three detection models by comparing the data of the five comparison models with the three detection models.

In this embodiment, three detection models and five comparison models are trained using a training set.

In this embodiment, as the number of model iterations increases, the loss curves of the three detection models are shown in Figures 9, 10 and 11, respectively. The scatter points in the figures show the loss performance of the model on the training set and the loss performance of the model on the validation set. Combining Figures 9 to 11, it can be seen that the model losses of the three detection models YOLOv5-L, YOLOv5-ML and YOLOv5-RML tend to saturate after about 100 training steps, among which the final loss value of YOLOv5 RML is 0.017.

In this embodiment, after the three detection models and five comparison models are trained, the model performance is verified on the verification set, and the results are shown in Table 2.

Table 2: Comparison of model validation results

Combined with Table 2, we can see that in terms of model accuracy, the three detection models are all higher than 0.9, and the five comparison models are all lower than 0.9; in terms of recall rate, the three detection models are all higher than 0.86, and the five comparison models are all lower than 0.84; in terms of average precision, the three detection models are all higher than 0.93, and the five comparison models are all lower than or equal to 0.92; in terms of F1 score, the three detection models are all higher than 0.89, and the five comparison models are all lower than 0.86; in terms of the number of parameters, the three detection models are all lower than 5.61, and the five comparison models are all higher than 7.02.

In particular, the third detection model YOLOv5-RML, compared with the original model YOLOv5, has an average accuracy improvement of 6.86%, a reduction of inference time of 4.82%, and a reduction of model weight of 18.23%; compared with YOLOv8, the accuracy, recall rate, average precision and F1 score have been improved by 3.28%, 8.46%, 3.79% and 5.89% respectively. It can be seen that the third detection model proposed in the present invention has better performance than YOLOv8 in the surface detection of precast concrete parts, which greatly promotes the technical progress of surface detection of precast concrete parts.

It can be seen that the three detection models proposed in the present invention have higher accuracy than the existing comparison models, and their convergence speed is also better than most of the existing comparison models. The comparison model YOLOv8 with a faster convergence speed has a large difference in accuracy from any detection model.

Referring to Figure 12, in this embodiment, the average accuracy (area under the P-R curve) of each model in Table 2 is further displayed through the P-R curve. The larger the area under the P-R curve, the better the model performance. It can be seen that the performance of the third detection model proposed in the present invention is better than that of YOLOv8.

In this embodiment, the detection accuracy of the trained detection model on different types of cracks is further verified on the validation set. For ease of representation, transverse cracks are marked with D00, longitudinal cracks are marked with D10, and fatigue cracks are marked with D20. The prediction accuracy of the detection model Yolov5-L for different types of crack images is shown in the confusion matrix of Figure 13, the prediction accuracy of the detection model Yolov5-ML for different types of crack images is shown in the confusion matrix of Figure 14, and the prediction accuracy of the detection model Yolov5-RML for different types of crack images is shown in the confusion matrix of Figure 15. Combined with the confusion matrix, it can be seen that compared with other cases, the third detection model YOLOv5 RML shows better performance in terms of accuracy, recall rate and mAP50 on longitudinal cracks, transverse cracks and fatigue cracks.

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.

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.

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

Claims (10)

1. A training method for a precast concrete crack detection model, comprising the following steps: First, a basic model and learning samples are obtained. The learning samples are the associated crack shooting images and crack annotated images. The basic model is based on the YOLOv5 model, and the connected Upsample unit and Concat unit are replaced by the LW module. The LW module first performs pooling on the input data of the Upsample unit in the YOLOv5 model, and then performs three-way splicing of the pooled data with the two-way input of the Upsample unit and Concat unit combination module in the YOLOv5 model and outputs the result. Then the basic model is made to learn the learning samples to iterate the model parameters until convergence; the converged basic model is used as a precast concrete crack detection model, whose input is the captured image and output is the crack annotation image. 2. The training method of the precast concrete crack detection model according to claim 1 is characterized in that the LW module includes an average pooling layer, a maximum pooling layer, a dimension superposition unit and a fifth Concat unit; the fifth Concat unit is provided with three input ends; the LW module is provided with two input ends; in the Upsample unit and the Concat unit combination module in the YOLOv5 model, the input end of the Upsample unit is recorded as the first input end of the combination module, and the input end of the Concat unit connected to the second C3 module is recorded as the second input end of the combination module; the first input end of the LW module replaces the first input end of the combination module, and the second input end of the LW module replaces the second input end of the combination module; The first input end of the LW module is respectively connected to the input end of the average pooling layer, the input end of the maximum pooling layer and the first input end of the fifth Concat unit, the output end of the average pooling layer and the output end of the maximum pooling layer are both connected to the input end of the dimension superposition unit, the output end of the dimension superposition unit is connected to the second input end of the fifth Concat unit, and the third input end of the fifth Concat unit is connected to the second input end of the LW module; the output end of the fifth Concat unit serves as the output end of the LW module. 3. The training method of the precast concrete crack detection model as described in claim 1 is characterized in that, in the YOLOv5 model, the modules with the same structure are named sequentially along the data flow direction; on the basis of the YOLOv5 model, the basic model also replaces the first C3 module and the fourth C3 module in the data flow direction with an MMT module, and the MMT module includes a Bottleneck unit, a MultiAttenCat unit and a tenth Conv unit connected in sequence; the input end of the MMT module is respectively connected to the input end of the Bottleneck unit and the input end of the MultiAttenCat unit, and the output end of the tenth Conv unit is used as the output end of the MMT module. 4. The training method of the precast concrete crack detection model as described in claim 3 is characterized in that the basic model is based on the YOLOv5 model, and the fourth Conv unit and the fifth Conv unit are replaced by RepVGG units. 5. The training method of the precast concrete crack detection model according to claim 1, wherein the iterative process of the model parameters comprises the following steps: St1, divide the learning samples into training set and test set; St2, extract multiple samples from the training set as training samples, and let the basic model learn the training samples to iterate the model parameters; St3, extract multiple samples from the test set as test samples, let the basic model predict the test samples and output crack annotation images; St4, calculate the loss of the basic model on the test sample to determine whether the basic model converges; if not, return to step St2; if yes, fix the basic model. 6. The training method for the precast concrete crack detection model as described in claim 5 is characterized in that, in St4, the loss of the basic model is a cross entropy loss or a mean square error loss. 7. A precast concrete crack detection method using the precast concrete crack detection model training method as described in any one of claims 1 to 6, characterized in that the detection model is first obtained by using the precast concrete crack detection model training method as described in any one of claims 1 to 6; then the target to be detected is photographed to obtain a target image; then the target image is input into the detection model, and the detection model outputs a crack annotation image to annotate the crack type. 8. The precast concrete crack detection method according to claim 7, wherein the crack types include: transverse cracks, longitudinal cracks and fatigue cracks. 9. A precast concrete crack detection 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 training method of the precast concrete crack detection model according to any one of claims 1 to 6. 10. A readable medium, characterized in that a computer program is stored therein, and when the computer program is executed, it is used to implement the training method of the precast concrete crack detection model according to any one of claims 1 to 6.