Concrete structure disease real-time visual perception method and handheld device

Abstract

The invention relates to the technical field of informatization and intelligent construction, and provides a concrete structure disease real-time visual perception method and handheld equipment. The handheld equipment comprises a data acquisition system and a data analysis system; the data acquisition system comprises a visual sensor, a miniature industrial personal computer, a laser displacement meter, a wireless transmission module, a low-power illuminating lamp and a power supply module; the data analysis system comprises a multi-type disease identification module and a fine crack high-precision quantification module; the multi-type disease identification module adopts a lightweight Faster-RTDETR deep learning model; the fine crack high-precision quantization module adopts a fine crack high-precision segmentation Half-AU2 model, and adopts an eight-direction search method to calculate crack width, length and area information. According to the method, the concrete structure diseases can be better and visually displayed.

Description

Technical Field

[0001] The present invention relates to the field of information technology and intelligent construction technology, and in particular to a real-time visual perception method for concrete structure defects and a handheld device. Background Art

[0002] With the rapid development of the economy and the gradual advancement of urbanization, the volume of bridge construction is expanding. During construction, concrete structures are prone to appearance quality defects such as exposed reinforcement, honeycombs, holes, slag inclusions, looseness, cracks, defects in connection parts, shape defects, and surface defects. The quality of concrete not only has a great impact on the safety of building structures, but also on the cost of construction projects. Therefore, concrete quality testing is an especially important part of the entire testing work. Traditional methods rely on professional inspectors and visual inspections, which are time-consuming and labor-intensive, with high testing costs and are greatly affected by human subjectivity. In recent years, with the successful application of deep learning models represented by convolutional neural networks in many computer vision fields, many deep learning-based defect detection methods have also been widely used in various industrial scenarios, but a lot of manpower and material resources are still required in data collection and subsequent processing. Therefore, there is an urgent need for a mobile perception and real-time detection method for multiple types of concrete structure defects to efficiently and accurately identify hidden defects. Summary of the invention

[0003] The present invention provides a real-time visual perception method for concrete structure defects and a handheld device, which have the advantages of non-contact, automation, high precision, and convenient carrying, and can be used for the detection and evaluation of concrete construction quality.

[0004] According to the present invention, a handheld device for real-time visual perception of concrete structure defects includes a data acquisition system and a data analysis system; the data acquisition system includes a visual sensor, a micro industrial computer, a laser displacement meter, a wireless transmission module, a low-power lighting lamp and a power module; the visual sensor is used to photograph the quality defects of the building structure and record the defect type and size information; the micro industrial computer is used to receive the image recorded by the visual sensor to achieve real-time detection of various types of quality defects; the laser displacement meter is used to determine the location information of the defect; the wireless transmission module is used for information transmission between the visual sensor and the micro industrial computer; and the power module is used to provide energy supply;

[0005] The data analysis system includes a multi-type disease recognition module and a fine crack high-precision quantification module; the multi-type disease recognition module adopts the lightweight Faster-RTDETR deep learning model, which includes the RepGELAN-VGG lightweight backbone network, CAA-FPN feature fusion module and AIFI-HiLo module; the fine crack high-precision quantification module adopts the fine crack high-precision segmentation Half-AU 2The model includes an image feature efficient extraction IRCGA module, a pooling HWD module and a Half-Net structure, and an eight-direction search method is used to calculate the crack width, length and area information.

[0006] As a preferred method, the RepGELAN-VGG lightweight backbone network is designed with the ELAN module, Repconv module and Ghostnet network ideas; for the input 640×640×3 RGB image, RepGELAN-VGG first changes the number of channels to 64 through a convolution operation of size 3×3 and stride 2, and then implements downsampling and feature extraction through 3×3 convolution layers and RepGELAN modules, and performs ReLu activation, and finally achieves 32 times downsampling of the original image.

[0007] Preferably, in the feature extraction part of the CAA-FPN feature fusion module, the feature map extracted by the RepGELAN-VGG lightweight backbone network is first input into the CAA module to further extract the contextual interdependence between distant pixels, and the channel matching dimension of the feature map is changed by 1×1 convolution; then feature fusion is performed. In order to further improve the model's ability to learn the correlation between feature maps at different levels, the high-level features are upsampled and expanded in size, and then the attention weights are generated by the CAA module, and multiplied with the low-level feature maps to filter the necessary semantic information in the low-level features, so as to realize the fusion of features of different scales and improve the model's feature learning ability.

[0008] As a preferred method, the HiLo attention mechanism is introduced in the AIFI-HiLo module, which has the ability to extract high- and low-frequency information and can consider the characteristics of different underlying frequencies.

[0009] Preferably, the image feature efficient extraction IRCGA module combines the meta-mobile module MMB and the cascade group attention CGA module; the image feature efficient extraction IRCGA module contains three branches, the first one is used for direct connection in the residual module, the second path first passes through a 1×1 convolution layer to change the number of channels, and then calculates with the attention weights generated by the CGA module in the third path to obtain the attention-weighted feature output, and then passes through a 3×3 convolution layer to extract local features, and performs shortcuts to alleviate the gradient vanishing problem in the deep network, and then performs a 1×1 convolution to change the number of channels, and finally splices with the direct connection feature map in the first branch to form the final output of the module.

[0010] Preferably, in the pooling HWD module, when the image enters the HWD module, it is first compressed to half of its original size through the HWT process, and at the same time, the number of channels is expanded to 4 times through four branches, and then the number of image channels is changed to the output size through a 3×3 or 1×1 convolutional layer, thereby achieving lossless compression of the feature map.

[0011] As a preference, in the Half-Net structure, retain U 2 Net’s Encode part, and optimized the Decoder module; first, the input image passes through the RSU7, RSU6, RSU5, RSU4, and RSU4F modules in turn to obtain the feature map of the corresponding level; in the Decoder part, a 1×1 convolution is used to change the number of channels and then directly output to obtain six significance prediction probability maps d6, d5, d4, d3, d2, and d1, and then these six feature maps are concat spliced, and finally the channels are changed through a 1×1 convolution and activated with a sigmoid function to obtain the final significance prediction probability map.

[0012] The present invention provides a method for real-time visual perception of concrete structure defects, which adopts the above-mentioned handheld device for real-time visual perception of concrete structure defects and includes the following steps:

[0013] 1) The data acquisition system takes images;

[0014] 2) Real-time detection of multiple types of diseases;

[0015] First, the input image passes through the RepGELAN-VGG lightweight backbone network, and then through 3×3 convolution and RepGELAN modules for convolution downsampling and efficient feature extraction, and the three feature layers P3, P4, and P5 are input into the CAA-FPN feature fusion network; then, in the CAA-FPN feature fusion module, each input feature layer first passes through the CAA attention module to strengthen the focus on contextual information, and after 1×1 convolution matching channels, it is multiplied with the adjacent feature layer to filter out the necessary semantic information in the low-level features, realizing the fusion of features of different scales; finally, the three feature layers after fusion representation are concat spliced ​​and input into the auxiliary prediction head decoder of the Decoder part to evaluate and output the probability of detecting the target;

[0016] 3) High-precision segmentation and quantification of concrete structure cracks;

[0017] First, the Half-AU2 segmentation model is trained on the constructed concrete structure fine crack dataset to obtain the optimal training weights; the image is handed over to the segmentation model for real-time recognition, and the segmentation results are returned; finally, the identified crack skeleton and edges are extracted, and the crack width, length and area are quantified.

[0018] Preferably, in step 3), specifically:

[0019] High-precision segmentation of Half-AU in fine cracks 2In the model, for the input image, the RSU module is first used to efficiently extract image features, and then one branch is used to output the process feature map after convolution, and the other branch is used to enhance the expression of image features through the IGCRA attention module. After repeating the above process four times, the feature map enters the RSU4F module, at which time features are no longer extracted and the image size is no longer changed. Only the number of channels is increased to reduce the loss of details on the small feature map. Finally, the process feature map of each branch is resized and spliced ​​with the channels for final output;

[0020] The crack mask image after semantic segmentation is used to extract its skeleton and edge with high precision to quantify the length and width of the crack. The Zhang-Suen algorithm is used to extract the crack skeleton, and the Sobel algorithm is used to extract the crack edge. Then, the crack width is calculated using the eight-direction search method, and the length is approximately equal to the crack skeleton length.

[0021] The segmentation result of the crack is a grayscale image containing only black and white. The area S of the crack can be approximately calculated using the differential element method. c,p , the formula is as follows:

[0022]

[0023] In the formula, ds represents the area element, which can be approximately regarded as the area of ​​a pixel; h and w represent the height and width of the image respectively; p(i,j) represents the normalized pixel value of the pixel at position (i,j), that is, when the pixel value of the point is 0, p(i,j) = 0, and when the pixel value of the point is 255, p(i,j) = 255 / 255 = 1; s represents the area of ​​a pixel;

[0024] The skeletonization result of the crack shows the topological structure and development law of the crack, so the length of the crack skeleton line can be approximately reflected as the length of the crack L c,p , the calculation formula is as follows:

[0025]

[0026] Among them, dl represents the length element, which is approximately the length of a pixel; l represents the length of a pixel point;

[0027] Then, the eight-direction search method is used to calculate the crack width. Starting from each pixel point on the crack skeleton, the search is carried out in eight directions: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The search is stopped when the edge of the crack is reached. Secondly, the Euclidean distance of the crack pixel point in each direction is counted, and the search distances in opposite directions are added to obtain the distances in four directions. Finally, the minimum value of the distances in the four directions is taken as the approximate crack pixel width at the pixel point. The crack width W c,p The calculation formula is as follows:

[0028] W i =d i +d 8-i

[0029] W c,p =min(W1,W2,W3,W4)

[0030] Where, d i Represents the distance from a point on the crack skeleton to the crack edge, i=1, 2, 3, 4.

[0031] The beneficial effects of the present invention are:

[0032] The invention is lightweight, portable and low-cost, not easily affected by factors such as space and environment, and has the advantages of good detection autonomy. The lightweight detection algorithm has higher recognition accuracy and target recognition confidence, meeting engineering needs. The crack segmentation and quantification method proposed in the invention has high quantification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of a handheld device for real-time visual perception of concrete structure diseases in an embodiment;

[0034] Figure 2 It is a schematic diagram of the lightweight Faster-RTDETR deep learning model in the embodiment;

[0035] Figure 3 It is the detection of multiple types of diseases of experimental bridge piers under complex background in the embodiment;

[0036] Figure 4 High-precision segmentation of fine cracks in the embodiment of Half-AU 2 Schematic diagram of the model;

[0037] Figure 5 This is a schematic diagram of the eight-direction search method for crack quantification in the embodiment;

[0038] Figure 6 It is a schematic diagram of the segmentation and quantification results of cracks in the actual bridge pier in the embodiment. DETAILED DESCRIPTION

[0039] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0040] Example

[0041] like Figure 1As shown, this embodiment provides a handheld device for real-time visual perception of concrete structure defects (the handheld device is specifically a handheld mobile detection device), which includes a data acquisition system and a data analysis system.

[0042] The data acquisition system includes (high-resolution) visual sensors, micro industrial computers, laser displacement meters, wireless transmission modules, low-power lighting and power modules; the visual sensors are used to capture quality defects of building structures and record defect types and size information; the micro industrial computers are used to receive images recorded by the visual sensors to achieve real-time detection of various types of quality defects; the laser displacement meter is used to determine the location information of the defects, and emits laser light to illuminate the object. After reflection, the receiver receives the signal and determines the distance based on the time it takes to propagate in the air; the wireless transmission module is used to transmit information between the visual sensor and the micro industrial computer; and the power module is used to provide energy supply.

[0043] The data analysis system includes a multi-type disease recognition module and a fine crack high-precision quantification module; the multi-type disease recognition module uses a lightweight Faster-RTDETR deep learning model, such as Figure 2 As shown in Figure 2, the model includes the RepGELAN-VGG lightweight backbone network, the CAA-FPN feature fusion module, and the AIFI-HiLo module; the fine crack high-precision quantization module uses the fine crack high-precision segmentation Half-AU 2 Models such as Figure 4 As shown, it includes the image feature efficient extraction IRCGA module, the pooling HWD module and the Half-Net structure, and the eight-direction search method is used to calculate the crack width, length and area information.

[0044] The RepGELAN-VGG lightweight backbone network is designed based on the ELAN module, Repconv module and Ghostnet network ideas. For the input 640×640×3 RGB image, RepGELAN-VGG first changes the number of channels to 64 through a convolution operation of size 3×3 and stride 2, and then implements downsampling and feature extraction through 3×3 convolutional layers and RepGELAN modules, and performs ReLu activation, and finally achieves 32 times downsampling of the original image.

[0045] In the feature extraction part of the CAA-FPN feature fusion module, the feature map extracted by the RepGELAN-VGG lightweight backbone network is first input into the CAA module to further extract the contextual interdependence between long-distance pixels, and the channel matching dimension of the feature map is changed through 1×1 convolution; then feature fusion is performed. In order to further improve the model's ability to learn the correlation between feature maps at different levels, the high-level features are upsampled and expanded in size, and then the attention weights are generated by the CAA module, and multiplied with the low-level feature maps to filter the necessary semantic information in the low-level features, so as to realize the fusion of features of different scales and improve the model's feature learning ability.

[0046] In the AIFI-HiLo module, the HiLo attention mechanism is introduced, which has the ability to extract high- and low-frequency information and can consider the characteristics of different underlying frequencies.

[0047] The IRCGA module for efficient image feature extraction combines the meta-mobile module MMB with the cascaded group attention CGA module; the IRCGA module for efficient image feature extraction contains three branches. The first one is used for direct connection in the residual module. The second path first changes the number of channels through a 1×1 convolution layer, and then calculates with the attention weights generated by the CGA module in the third path to obtain the attention-weighted feature output, and then extracts local features through a 3×3 convolution layer, and performs shortcuts to alleviate the gradient vanishing problem in the deep network. After that, a 1×1 convolution is performed to change the number of channels, and finally it is spliced ​​with the direct connection feature map in the first branch to form the final output of the module.

[0048] In the pooling HWD module, when the image enters the HWD module, it first undergoes the HWT process to compress the image size to half of its original size, and at the same time, it expands the number of channels to 4 times through four branches, and then passes through a 3×3 or 1×1 convolutional layer to change the number of image channels to the output size, thereby achieving lossless compression of the feature map.

[0049] In the Half-Net structure, U is retained 2 Net’s Encode part, and optimized the Decoder module; first, the input image passes through the RSU7, RSU6, RSU5, RSU4, and RSU4F modules in turn to obtain the feature map of the corresponding level; in the Decoder part, a 1×1 convolution is used to change the number of channels and then directly output to obtain six significance prediction probability maps d6, d5, d4, d3, d2, and d1, and then these six feature maps are concat spliced, and finally the channels are changed through a 1×1 convolution and activated with a sigmoid function to obtain the final significance prediction probability map.

[0050] The present invention provides a method for real-time visual perception of concrete structure defects, which adopts the above-mentioned handheld device for real-time visual perception of concrete structure defects and includes the following steps:

[0051] 1) The data acquisition system takes images;

[0052] Hold the device and adjust different postures (standing upright, squatting, looking down, looking up, etc.) to achieve multi-directional and multi-angle detection of the target area to improve detection efficiency and accuracy.

[0053] 2) Real-time detection of multiple types of diseases;

[0054] First, the input image passes through the RepGELAN-VGG lightweight backbone network, and then through 3×3 convolution and RepGELAN modules for convolution downsampling and efficient feature extraction, and the three feature layers P3, P4, and P5 are input into the CAA-FPN feature fusion network; then, in the CAA-FPN feature fusion module, each input feature layer first passes through the CAA attention module to strengthen the attention to contextual information, and after 1×1 convolution matching channels, it is multiplied with the adjacent feature layer to filter out the necessary semantic information in the low-level features, so as to realize the fusion of features of different scales; finally, the three feature layers after fusion representation are concat spliced ​​and input into the auxiliary prediction head decoder of the Decoder part to evaluate and output the probability of detecting the target. Figure 3 To detect multiple types of diseases in experimental bridge piers under complex backgrounds.

[0055] 3) High-precision segmentation and quantification of concrete structure cracks;

[0056] First, the Half-AU2 segmentation model is trained on the constructed concrete structure fine crack dataset to obtain the optimal training weights; the image is handed over to the segmentation model for real-time recognition, and the segmentation results are returned; finally, the identified crack skeleton and edges are extracted, and the crack width, length and area are quantified.

[0057] In step 3), specifically:

[0058] High-precision segmentation of Half-AU in fine cracks 2 In the model, for the input image, the RSU module is first used to efficiently extract image features, and then one branch is used to output the process feature map after convolution, and the other branch is used to enhance the expression of image features through the IGCRA attention module. After repeating the above process four times, the feature map enters the RSU4F module, at which time features are no longer extracted and the image size is no longer changed. Only the number of channels is increased to reduce the loss of details on the small feature map. Finally, the process feature map of each branch is resized and spliced ​​with the channels for final output;

[0059] The crack mask image after semantic segmentation is used to extract its skeleton and edge with high precision to quantify the length and width of the crack. The Zhang-Suen algorithm is used to extract the crack skeleton, and the Sobel algorithm is used to extract the crack edge. Then, the crack width is calculated using the eight-direction search method, and the length is approximately equal to the crack skeleton length.

[0060] The segmentation result of the crack is a grayscale image containing only black and white (pixel values ​​are 0 and 255 respectively). The area S of the crack can be approximately calculated using the differential element method. c,p , the formula is as follows:

[0061]

[0062] In the formula, ds represents the area element, which can be approximately regarded as the area of ​​a pixel; h and w represent the height and width of the image respectively; p(i,j) represents the normalized pixel value of the pixel at position (i,j), that is, when the pixel value of the point is 0, p(i,j) = 0, and when the pixel value of the point is 255, p(i,j) = 255 / 255 = 1; s represents the area of ​​a pixel;

[0063] The skeletonization result of the crack shows the topological structure and development law of the crack, so the length of the crack skeleton line can be approximately reflected as the length of the crack L c,p , the calculation formula is as follows:

[0064]

[0065] Wherein, dl represents a length element, which is approximately considered to be the length of a pixel (in this embodiment, a pixel is considered to be a square, so the pixel length and width are the same); l represents the length of a pixel point;

[0066] Then, if Figure 5 As shown in the figure, the eight-direction search method is used to calculate the crack width. Starting from each pixel point on the crack skeleton, the search is carried out in eight directions of 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° respectively. The search is stopped when the edge of the crack is reached. Secondly, the Euclidean distance of the crack pixel point in each direction is counted, and the search distances in opposite directions are added to obtain the distances in four directions. Finally, the minimum value of the distances in the four directions is taken as the approximate crack pixel width at the pixel point. The crack width W c,p The calculation formula is as follows:

[0067] W i =d i +d 8-i

[0068] W c,p =min(W1,W2,W3,W4)

[0069] Where, d i Represents the distance from a point on the crack skeleton to the crack edge, i=1, 2, 3, 4.

[0070] Figure 6 The results of the segmentation and quantification of cracks in the bridge piers of a real bridge are shown in the figure. It can be seen that the model accurately segments the cracks in the bridge piers in the figure, with few false detections and missed detections.

[0071] The present invention and its embodiments are described schematically above, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and creatively designs a structure and an embodiment similar to the technical solution without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A handheld device for real-time visual perception of concrete structure defects, characterized by: It includes a data acquisition system and a data analysis system; the data acquisition system includes a visual sensor, a micro industrial computer, a laser displacement meter, a wireless transmission module, a low-power lighting lamp and a power module; the visual sensor is used to photograph the quality defects of the building structure and record the defect type and size information; the micro industrial computer is used to receive the images recorded by the visual sensor to realize real-time detection of various types of quality defects; the laser displacement meter is used to determine the location information of the defect; the wireless transmission module is used for information transmission between the visual sensor and the micro industrial computer; the power module is used to provide energy supply; The data analysis system includes a multi-type disease recognition module and a fine crack high-precision quantification module; the multi-type disease recognition module adopts the lightweight Faster-RTDETR deep learning model, which includes the RepGELAN-VGG lightweight backbone network, CAA-FPN feature fusion module and AIFI-HiLo module; the fine crack high-precision quantification module adopts the fine crack high-precision segmentation Half-AU 2 The model includes an image feature efficient extraction IRCGA module, a pooling HWD module and a Half-Net structure, and an eight-direction search method is used to calculate the crack width, length and area information.

2. A handheld device for real-time visual perception of concrete structure defects according to claim 1, characterized in that: The RepGELAN-VGG lightweight backbone network is designed based on the ELAN module, Repconv module and Ghostnet network ideas. For the input 640×640×3 RGB image, RepGELAN-VGG first changes the number of channels to 64 through a convolution operation of size 3×3 and stride 2, and then implements downsampling and feature extraction through 3×3 convolutional layers and RepGELAN modules, and performs ReLu activation, and finally achieves 32 times downsampling of the original image.

3. The handheld device for real-time visual perception of concrete structure defects according to claim 2 is characterized by: In the feature extraction part of the CAA-FPN feature fusion module, the feature map extracted by the RepGELAN-VGG lightweight backbone network is first input into the CAA module to further extract the contextual interdependence between long-distance pixels, and the channel matching dimension of the feature map is changed through 1×1 convolution; then feature fusion is performed. In order to further improve the model's ability to learn the correlation between feature maps at different levels, the high-level features are upsampled and expanded in size, and then the attention weights are generated by the CAA module, and multiplied with the low-level feature maps to filter the necessary semantic information in the low-level features, so as to realize the fusion of features of different scales and improve the model's feature learning ability.

4. The handheld device for real-time visual perception of concrete structure defects according to claim 3 is characterized by: In the AIFI-HiLo module, the HiLo attention mechanism is introduced, which has the ability to extract high- and low-frequency information and can consider the characteristics of different underlying frequencies.

5. A handheld device for real-time visual perception of concrete structure defects according to claim 4, characterized in that: The IRCGA module for efficient image feature extraction combines the meta-mobile module MMB with the cascaded group attention CGA module; the IRCGA module for efficient image feature extraction contains three branches. The first one is used for direct connection in the residual module. The second path first changes the number of channels through a 1×1 convolution layer, and then calculates with the attention weights generated by the CGA module in the third path to obtain the attention-weighted feature output, and then extracts local features through a 3×3 convolution layer, and performs shortcuts to alleviate the gradient vanishing problem in the deep network. After that, a 1×1 convolution is performed to change the number of channels, and finally it is spliced ​​with the direct connection feature map in the first branch to form the final output of the module.

6. The handheld device for real-time visual perception of concrete structure defects according to claim 5 is characterized by: In the pooling HWD module, when the image enters the HWD module, it first undergoes the HWT process to compress the image size to half of its original size, and at the same time, it expands the number of channels to 4 times through four branches, and then passes through a 3×3 or 1×1 convolutional layer to change the number of image channels to the output size, thereby achieving lossless compression of the feature map.

7. A handheld device for real-time visual perception of concrete structure defects according to claim 6, characterized in that: In the Half-Net structure, U is retained 2 Net’s Encode part, and optimized the Decoder module; first, the input image passes through the RSU7, RSU6, RSU5, RSU4, and RSU4F modules in turn to obtain the feature map of the corresponding level; in the Decoder part, a 1×1 convolution is used to change the number of channels and then directly output to obtain six significance prediction probability maps d6, d5, d4, d3, d2, and d1, and then these six feature maps are concat spliced, and finally the channels are changed through a 1×1 convolution and activated with a sigmoid function to obtain the final significance prediction probability map.

8. A real-time visual perception method for concrete structure defects, characterized by: It uses a handheld device for real-time visual perception of concrete structure defects as described in claim 7, and includes the following steps: 1) The data acquisition system takes images; 2) Real-time detection of multiple types of diseases; First, the input image passes through the RepGELAN-VGG lightweight backbone network, and then through 3×3 convolution and RepGELAN modules for convolution downsampling and efficient feature extraction, and the three feature layers P3, P4, and P5 are input into the CAA-FPN feature fusion network; then, in the CAA-FPN feature fusion module, each input feature layer first passes through the CAA attention module to strengthen the focus on contextual information, and after 1×1 convolution matching channels, it is multiplied with the adjacent feature layer to filter out the necessary semantic information in the low-level features, realizing the fusion of features of different scales; finally, the three feature layers after fusion representation are concat spliced ​​and input into the auxiliary prediction head decoder of the Decoder part to evaluate and output the probability of detecting the target; 3) High-precision segmentation and quantification of concrete structure cracks; First, the Half-AU2 segmentation model is trained on the constructed concrete structure fine crack dataset to obtain the optimal training weights; the image is handed over to the segmentation model for real-time recognition, and the segmentation result is returned; finally, the identified crack skeleton and edge are extracted, and the crack width, length and area are quantified.

9. A real-time visual perception method for concrete structure defects according to claim 8, characterized in that: In step 3), specifically: High-precision segmentation of Half-AU in fine cracks 2 In the model, for the input image, the RSU module is first used to efficiently extract image features, and then one branch is used to output the process feature map after convolution, and the other branch is used to enhance the expression of image features through the IGCRA attention module; after repeating the above process four times, the feature map enters the RSU4F module, at this time, no features are extracted and the image size is no longer changed, but only the number of channels is increased to reduce the loss of details on the small feature map; finally, the process feature map of each branch is resized and spliced ​​with the channel for final output; The crack mask image after semantic segmentation is used to extract its skeleton and edge with high precision to quantify the length and width of the crack. The Zhang-Suen algorithm is used to extract the crack skeleton, and the Sobel algorithm is used to extract the crack edge. Then, the crack width is calculated using the eight-direction search method, and the length is approximately equal to the crack skeleton length. The result of crack segmentation is a grayscale image containing only black and white. The area S of the crack can be approximately calculated using the differential element method. c,p , the formula is as follows: In the formula, ds represents the area element, which can be approximately regarded as the area of ​​a pixel; h and w represent the height and width of the image respectively; p(i,j) represents the normalized pixel value of the pixel at position (i,j), that is, when the pixel value of the point is 0, p(i,j) = 0, and when the pixel value of the point is 255, p(i,j) = 255 / 255 = 1; s represents the area of ​​a pixel; The skeletonization result of the crack shows the topological structure and development law of the crack, so the length of the crack skeleton line can be approximately reflected as the length of the crack L c,p , the calculation formula is as follows: Among them, dl represents the length element, which is approximately the length of a pixel; l represents the length of a pixel point; Then, the eight-direction search method is used to calculate the crack width. Starting from each pixel point on the crack skeleton, the search is carried out in eight directions: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The search is stopped when the edge of the crack is reached. Secondly, the Euclidean distance of the crack pixel point in each direction is counted, and the search distances in opposite directions are added to obtain the distances in four directions. Finally, the minimum value of the distances in the four directions is taken as the approximate crack pixel width at the pixel point. The crack width W c,p The calculation formula is as follows: W i =d i +d 8-i <h2 style=";text-align:left;direction:ltr">W<h2 style=";text-align:left;direction:ltr"> c,p <h2 style=";text-align:left;direction:ltr"> =min(W1,W2,W3,W4) Where, d i Represents the distance from a point on the crack skeleton to the crack edge, i=1, 2, 3, 4.

Images