Intelligent tunnel disease detection and identification method and system based on visual perception

Abstract

The invention provides a tunnel disease intelligent detection and identification method and system based on visual perception, and relates to the technical field of tunnel disease detection. Comprising the steps of obtaining a tunnel disease picture data set; a DB-FOCS network model is built; the DB-FOCS network model is trained, and a trained DB-FOCS network model is obtained; inputting a to-be-identified tunnel disease picture into the DB-FOCS network model, extracting features by using a deformable convolutional layer and generating an offset, and correcting the extracted features by using the offset to obtain a multi-layer feature map; inputting the multi-layer feature map into a BiFPN structure, and carrying out cross-layer feature fusion; and inputting the cross-layer feature map into a Head branch, and extracting features by using the deformable convolutional layer again to obtain a disease identification result. According to the method, the algorithm can better cope with various size changes and deformation in a target detection task, the target detection performance is improved, and the model accuracy and stability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel disease detection, and in particular to a method and system for intelligent detection and identification of tunnel diseases based on visual perception. Background Art

[0002] Tunnels are an indispensable and important part of modern transportation construction, and their safe and stable operation is an important task to ensure smooth road traffic and the safety of people's lives and property. However, in the long-term use of tunnels, due to the influence of various factors, such as geological structure, climatic conditions, and traffic volume, various diseases such as water seepage, cracking, and collapse will occur. The existence of these diseases will not only threaten the safe and stable operation of the tunnel, but also affect the normal passage of traffic and bring inconvenience to people's travel. Therefore, how to identify tunnel diseases early and accurately is crucial to ensure the safe and stable operation of the tunnel and the safety of people's travel.

[0003] At present, the detection and identification of tunnel diseases mainly rely on two methods: manual inspection and instrument-assisted detection. Although manual inspection can effectively observe and record tunnel diseases, its low efficiency, high cost, and susceptibility to subjective factors restrict its application scope and effect. Although instrument-assisted detection can improve detection efficiency and accuracy, the equipment itself is expensive, and it is difficult to accurately detect and identify some hidden diseases, so it also has certain limitations.

[0004] Computer vision technology has gradually become a new and effective means of tunnel disease detection and identification. By using computer vision technology to analyze and process images inside the tunnel, it can automatically identify and judge the type and degree of diseases inside the tunnel. It can not only improve detection efficiency and accuracy, but also effectively avoid the influence of human factors and reduce detection costs.

[0005] However, the inventors found that the prior art still has the following technical problems:

[0006] (1) Based on the different shooting angles and distances of images in the tunnel, when applying computer vision technology to extract features, if the size changes and various deformations of the objects cannot be well dealt with, the features of tunnel defects cannot be accurately extracted, which will affect the accuracy of tunnel defect recognition in the later stage;

[0007] (2) Due to the particularity of the tunnel environment, the boundaries of the collected images are often fuzzy. When extracting target features with complex features and fuzzy boundaries, if the target image feature extraction capability and multi-scale target adaptability are not high, the recognition accuracy of tunnel hazards will be affected.

[0008] (3) When detecting a target, the loss function of the model will affect the positioning accuracy of the target box and the overall convergence speed of the model. The accuracy and confidence of the target box are often ignored in the existing technology, making it difficult to more accurately evaluate the performance of the target detection model. Summary of the invention

[0009] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a method and system for intelligent detection and identification of tunnel defects based on visual perception, introduces a deformable convolution layer in the FCOS model, improves the feature extraction capability of the backbone network, and enables the algorithm to better cope with various size changes and deformations in target detection tasks; introduces a BiFPN module, which can dynamically adjust the weights and proportions of the feature maps to better adapt to different target scales and shapes, thereby improving the performance of target detection; uses the EIoU loss function to improve the accuracy and stability of the model.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention provides a method for intelligent detection and identification of tunnel defects based on visual perception.

[0012] The intelligent detection and identification method of tunnel defects based on visual perception includes the following steps:

[0013] Obtain a dataset of tunnel disease images and annotate the images;

[0014] Based on the FOCS model, deformable convolutional layers are used to replace the third and fourth convolutions of the FOCS model backbone network and the first four convolutions of the two Head branches. The BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model.

[0015] The DB-FOCS network model is trained using the annotated tunnel disease images to obtain a trained DB-FOCS network model;

[0016] The tunnel disease image to be identified is input into the trained DB-FOCS network model, and the deformable convolution layer is used to extract features and generate deformed offsets. The deformed offsets are used to correct the extracted features and obtain multi-layer feature maps. The multi-layer feature maps are input into the BiFPN structure, and cross-layer feature fusion is performed to obtain cross-layer feature maps. The cross-layer feature maps are input into the Head branch, and the deformable convolution layer is used again to extract features to finally obtain the disease recognition results.

[0017] The second aspect of the present invention provides an intelligent tunnel disease detection and identification system based on visual perception.

[0018] The intelligent tunnel disease detection and identification system based on visual perception includes:

[0019] The data set acquisition module is configured to: acquire a tunnel disease image data set and annotate the images;

[0020] The model building module is configured as follows: based on the FOCS model, the deformable convolutional layers are used to replace the third and fourth convolutions of the backbone network of the FOCS model, and the first four convolutions of the two Head branches, and the BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model;

[0021] The training module is configured to: train the DB-FOCS network model using the annotated tunnel disease images to obtain a trained DB-FOCS network model;

[0022] The disease recognition module is configured as follows: input the tunnel disease image to be identified into the trained DB-FOCS network model, use the deformable convolution layer to extract features and generate deformed offsets, use the deformed offsets to correct the extracted features and obtain a multi-layer feature map; input the multi-layer feature map into the BiFPN structure, perform cross-layer feature fusion, and obtain a cross-layer feature map; input the cross-layer feature map into the Head branch, use the deformable convolution layer to extract features again, and finally obtain the disease recognition result.

[0023] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the method for intelligent detection and identification of tunnel defects based on visual perception as described in the first aspect of the present invention.

[0024] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for intelligent detection and identification of tunnel defects based on visual perception as described in the first aspect of the present invention are implemented.

[0025] One or more of the above technical solutions have the following beneficial effects:

[0026] The present invention proposes a method and system for intelligent detection and identification of tunnel defects based on visual perception, uses the FCOS model with good performance for deep learning, improves the regression accuracy by improving the border position regression loss function and using the EIOU loss function, and improves the network structure of the FCOS algorithm by introducing the highly portable BiFPN module and deformable convolution layer. Without affecting its convenience, its ability to extract features of target images with complex features and fuzzy boundaries and its multi-scale target adaptation ability are enhanced, thereby improving the recognition accuracy of tunnel disasters.

[0027] The present invention conducted a large number of experiments on public data sets, analyzed the effectiveness and superiority of the algorithm, and after comparative testing with the original algorithm, using precision, recall, PR curve, etc. as evaluation indicators, it was found that the improved algorithm has stronger overall detection accuracy and robustness.

[0028] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0030] Figure 1 This is a structural diagram of the FOCS model in the prior art.

[0031] Figure 2 This is a structural diagram of the deformable convolutional layer of the first embodiment.

[0032] Figure 3 This is the BiFPN structure diagram of the first embodiment.

[0033] Figure 4 This is a flow chart of the overall method of the first embodiment.

[0034] Figure 5 This is an example diagram of the verification results of the first embodiment. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0037] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0038] The overall idea proposed by the present invention is:

[0039] The intelligent detection and identification method for tunnel defects based on visual perception provided by the present invention realizes accurate identification and classification of tunnel defects by establishing an effective algorithm model and training data set. It has the advantages of high detection efficiency, high precision, low cost, strong adaptability, etc., and has broad application prospects in the field of tunnel defect detection and identification.

[0040] The specific steps include:

[0041] S1. Construct tunnel disease dataset;

[0042] S2. In the FOCS model, an improved deformable convolution layer (DConv) is used to replace the third and fourth stage convolutions of the backbone network and the first four convolutions of the two Head branches;

[0043] S3. In the FOCS model, the BiFPN structure is used to replace the FPN structure in the original model;

[0044] S4. In the FOCS model, the EIoU loss function is used to replace the IoU loss function in the original model;

[0045] S5, build a new DB-FCOS network structure;

[0046] S6, using the DB-FCOS network model constructed in S5, trains and optimizes the sample library to obtain the optimal model;

[0047] S7. Use the optimal model obtained in S6 to perform target detection and recognition on the validation set to obtain recognition results.

[0048] The present invention:

[0049] (1) A dataset of tunnel secondary lining defects during construction was constructed, and the FCOS model was introduced. The improved deformable convolutional layer was used to replace the third and fourth convolutions of the backbone network and the first four traditional convolutions of the two head branches, respectively. This improved the feature extraction capability of the backbone network and adaptively changed the shape and size of the convolution kernel according to the shape change of the target object, so as to more accurately extract the features of the target object.

[0050] (2) Use the BiFPN structure to replace the FPN structure and fuse feature maps from different levels to better capture the spatial and semantic information of the target;

[0051] (3) The EIoU loss function is used to replace the IoU loss function in the original model to improve the accuracy and stability of the model.

[0052] Embodiment 1

[0053] This embodiment discloses a method for intelligent detection and identification of tunnel defects based on visual perception.

[0054] like Figure 4 As shown, the intelligent detection and identification method of tunnel defects based on visual perception includes the following steps:

[0055] Obtain a dataset of tunnel disease images and annotate the images;

[0056] Based on the FOCS model, deformable convolutional layers are used to replace the third and fourth convolutions of the FOCS model backbone network and the first four convolutions of the two Head branches. The BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model.

[0057] The DB-FOCS network model is trained using the annotated tunnel disease images to obtain a trained DB-FOCS network model;

[0058] The tunnel disease image to be identified is input into the trained DB-FOCS network model, and the deformable convolution layer is used to extract features and generate deformed offsets. The deformed offsets are used to correct the extracted features and obtain multi-layer feature maps. The multi-layer feature maps are input into the BiFPN structure, and cross-layer feature fusion is performed to obtain cross-layer feature maps. The cross-layer feature maps are input into the Head branch, and the deformable convolution layer is used again to extract features to finally obtain the disease recognition results.

[0059] Aiming at the problem of tunnel secondary lining defects during construction, the present invention introduces a deformable convolutional layer into the FCOS model to improve the feature extraction capability of the backbone network, so that the algorithm can better cope with various size changes and deformations in target detection tasks; the BiFPN module can dynamically adjust the weights and proportions of the feature maps to better adapt to different target scales and shapes, thereby improving the performance of target detection; the use of the EIoU loss function can improve the accuracy and stability of the model.

[0060] To achieve the above object, the present invention adopts the following technical solutions.

[0061] S1. Construct a tunnel disease sample set. Use the pictures collected at the construction site and the typical pictures collected on the Internet to construct a tunnel disease dataset. Then use LableImg to annotate the samples, complete the construction of the sample library, and use the ImgAug library to enhance the sample set data.

[0062] S2, such as Figure 1 , Figure 2As shown in the figure, in the FOCS model, an improved deformable convolution layer (DConv) is used to replace the third and fourth stage convolutions of the backbone network and the first four convolutions of the two Head branches. ResNet-50 is used for feature extraction, in which the convolution module used is a conventional convolution module. Due to the different angles and distances of the images taken in the tunnel, it may not be able to cope with the size changes and various deformations of the object. The role of the deformable convolution layer is to improve the adaptability of the network to the shape and size changes of the target object. It can improve the feature extraction ability of the backbone network, adaptively change the shape and size of the convolution kernel according to the shape changes of the target object, and thus more accurately extract the features of the target object.

[0063] In general, the specific operation process of the deformable convolution layer is:

[0064] (1) Based on the input image, the feature map is extracted using the traditional convolution kernel.

[0065] (2) Take the obtained feature map as input and apply another convolution layer to the feature map. The purpose of this is to obtain the offset of the deformable convolution. The offset layer is 2N because we need to change both the x and y values ​​when translating on the plane.

[0066] (3) During training, the convolution kernel used to generate the output features and the convolution kernel used to generate the offset are learned synchronously. The offset is learned by using an interpolation algorithm and back propagation.

[0067] Generate offsets: The input features are operated through the convolution kernel used to generate offsets to obtain offsets. These offsets represent the spatial offset of each pixel, that is, how the features at each position need to be displaced in space to adapt to the shape changes of the target object.

[0068] Apply offset: Apply the generated offset to the input feature. This process actually transforms the input feature spatially so that the feature map can adaptively adapt to the shape and size changes of the target object. Adaptive deformation is achieved by offsetting each pixel accordingly.

[0069] Feature fusion: The input features with the applied offset are fused with the generated output features. Usually, this fusion can be done by weighted summation or concatenation to combine the original features with the adaptively deformed features, thereby obtaining a feature representation that is more adaptable to the shape changes of the target object.

[0070] S3, such as Figure 1 , Figure 3As shown in the figure, in the FOCS model, the BiFPN structure is used to replace the FPN structure in the original model. BiFPN is a weighted bidirectional feature pyramid network that can perform multi-scale feature fusion simply and quickly. The BiFPN structure consists of multiple BiFPN modules, each of which contains four steps: upsampling, downsampling, cross-layer feature fusion, and an adjustable attention mechanism.

[0071] The specific structural process is as follows:

[0072] (1) Input features: The input of BiFPN is feature maps from different layers. These feature maps are first processed by the backbone network and then sent to the BiFPN module.

[0073] (2) Upsampling: First, the lower-resolution feature map is increased in resolution through an upsampling operation to make it the same size as the higher-resolution feature map.

[0074] (3) Downsampling: The higher-resolution feature map is then reduced in resolution through a downsampling operation to make it the same size as the lower-resolution feature map.

[0075] (4) Cross-layer feature fusion: Next, the upsampled and downsampled feature maps are fused. Cross-layer feature fusion is used here, that is, feature maps from different layers are fused. This can better combine semantic information from different layers.

[0076] The fused multi-layer features are feature maps from different levels. In the task of object detection, multi-layer features extracted by backbone networks (such as ResNet) are usually used for object detection. These multi-layer features have different resolutions and semantic information. For example, shallow features may be more suitable for detecting small-sized objects, while deep features may contain richer semantic information.

[0077] (5) Adjustable attention mechanism: Finally, in order to further improve the quality of the fused feature map, BiFPN also uses an adjustable attention mechanism. Here, the attention mechanism in the SENet structure is adopted, that is, the information flow in the feature map is controlled by learning adjustable weights to further improve the quality of the feature map.

[0078] The above four steps will be repeated many times until the target detection result is finally output. The BiFPN module can fuse feature maps from different levels to utilize feature information at different levels to better capture the spatial and semantic information of the target.

[0079] S4. In the FOCS model, the EIoU loss function is used to replace the IoU loss function in the original model. The EIoU loss function is a loss function used for target detection tasks. Compared with the traditional IoU loss function, the EIoU loss function takes into account the accuracy and confidence of the target box, and can more accurately evaluate the performance of the target detection model.

[0080]

[0081] The loss function consists of three parts: overlap loss, center distance loss, and width and height loss. The first two parts continue the method in CIOU, but the width and height loss directly minimizes the difference between the width and height of the target box and the anchor box, making the convergence speed faster. h and C w It is the width and height of the smallest bounding box covering the two boxes.

[0082] Specifically, the EIoU loss function first calculates the IoU value of the predicted box and the true box, then weights it according to factors such as the distance between the center points of the predicted box and the true box, the length and width ratio, etc., to obtain a correction coefficient, and finally multiplies the IoU value by the correction coefficient to obtain the EIoU value. When training the tunnel disease model, using the EIoU loss function can improve the accuracy and stability of the model.

[0083] Among them, L IOU It is the ordinary intersection-over-union loss; the intersection-over-union loss measures the ratio of the intersection area of ​​two boxes to the union area. The larger the value, the higher the degree of overlap. dis Represents the distance loss between boxes, which is used to measure the distance between the predicted box and the center of the real box; L asp represents the aspect ratio loss, which is used to measure the difference in aspect ratio between the predicted box and the real box; ρ represents a weight factor used to adjust the impact of different loss items on the total loss; (b,b gt ) represents the position of the predicted box and the true box, usually a bounding box expressed in coordinate form; (w,w gt ) represents the width of the predicted box and the real box, (h,h gt ) represents the height of the predicted box and the true box, c is the diagonal length of the smallest enclosing box covering the two boxes, C h and C w It is the scaling factor of the width and height loss, which is used to adjust the magnitude of the loss so that boxes of different scales can be reasonably balanced in the loss calculation.

[0084] S5, build a new DB-FOCS network structure. The original structure is adjusted through S2, S3 and S4. The structure is mainly composed of three parts: Backbone network, BiFPN and Head output end. The structure mainly includes:

[0085] The backbone network structure used by FCOS is ResNet-50, which performs multi-layer convolution and activation function operations on the feature map to extract feature information.

[0086] The head structure of FCOS consists of three submodules: classification submodule, regression submodule and centrality submodule.

[0087] The classification submodule is a convolutional layer used to predict the category of the target for each feature point;

[0088] The regression submodule is also a convolutional layer, which is used to predict the bounding box offset and size of the target for each feature point;

[0089] The center degree submodule is also a convolutional layer, which is used to predict the center degree of the target for each feature point.

[0090] The head structure of FCOS is different from traditional object detection algorithms. It is based on a fully convolutional network, so it can process input images of any scale and does not require pre-setting anchor boxes, which improves detection performance.

[0091] S6. Use the DB-FCOS network model constructed in S5 to train and optimize the sample library to obtain the optimal model.

[0092] S7. Use the optimal model obtained in S6 to perform target detection and recognition on the validation set to obtain recognition results.

[0093] Tunnel projects are usually too large and complex, with massive amounts of data and too many uncertain factors, so practical problems such as insufficient equipment and insufficient personnel experience can easily lead to frequent accidents. To address these thorny issues, the present invention introduces deformable convolutional layers and BiFPN into the FCOS network, uses the EIOU loss function, and constructs a DB-FCOS network, which further enhances the network's multi-scale adaptive capabilities and detection accuracy, and improves the target detection capability for tunnel diseases. It has high practical value. In addition, the present invention can also be migrated to surface crack detection in other fields, and has strong applicability.

[0094] More specifically, this embodiment includes the following steps:

[0095] S1. Establish a sample set of tunnel secondary lining defects during the construction period.

[0096] This example uses pictures collected at the construction site and typical pictures collected on the Internet to construct a tunnel disease dataset, which has the characteristics of large scale changes and complex features. The resolution of each image is 680*680 pixels, the number of target categories is set to 2, the category names are set to crack and water, and a total of 4,000 training sample libraries and 800 test sample libraries are established.

[0097] S2. Experimental process design.

[0098] This example is based on the Pytorch deep learning framework, the integrated development and testing tool uses pycharm, the GPU hardware environment is GTX1080Ti, and the DB-FCOS deep learning network model designed above is used to train and test the samples. The final verification result example is as follows Figure 5 shown.

[0099] The main steps include:

[0100] First, the original FCOS network is used to train and test the samples;

[0101] Second, the DB-FCOS network proposed in the present invention is used to train and test the same samples;

[0102] Third, the detection accuracy of the two networks for the same samples is counted separately.

[0103] In this embodiment, the same network parameters are selected, epochs is set to 300, batch-size is 32, and img-size is [680 680]. Except for the data set and target category, other parameters remain unchanged by the default parameters of FCOS.

[0104] The following is the detailed processing process of the DB-FOCS network model on the input image:

[0105]

[0106]

[0107] Note:

[0108] (1) The “-1” in the previous layer column indicates that it comes from the previous layer;

[0109] (2) The Detect output layer [3, 80, 80, 85] indicates that the output size of the model is 80x80 pixels, and the model predicts 3 bounding boxes, each containing 80 feature maps and 85 channels.

[0110] Embodiment 2

[0111] This embodiment discloses a tunnel disease intelligent detection and identification system based on visual perception.

[0112] The intelligent tunnel disease detection and identification system based on visual perception includes:

[0113] The data set acquisition module is configured to: acquire a tunnel disease image data set and annotate the images;

[0114] The model building module is configured as follows: based on the FOCS model, the deformable convolutional layers are used to replace the third and fourth convolutions of the backbone network of the FOCS model, and the first four convolutions of the two Head branches, and the BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model;

[0115] The training module is configured to: train the DB-FOCS network model using the annotated tunnel disease images to obtain a trained DB-FOCS network model;

[0116] The disease recognition module is configured as follows: input the tunnel disease image to be identified into the trained DB-FOCS network model, use the deformable convolution layer to extract features and generate deformed offsets, use the deformed offsets to correct the extracted features and obtain a multi-layer feature map; input the multi-layer feature map into the BiFPN structure, perform cross-layer feature fusion, and obtain a cross-layer feature map; input the cross-layer feature map into the Head branch, use the deformable convolution layer to extract features again, and finally obtain the disease recognition result.

[0117] Embodiment 3

[0118] The purpose of this embodiment is to provide a computer-readable storage medium.

[0119] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for intelligent detection and identification of tunnel defects based on visual perception as described in Example 1 of the present disclosure.

[0120] Embodiment 4

[0121] The purpose of this embodiment is to provide an electronic device.

[0122] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for intelligent detection and identification of tunnel defects based on visual perception as described in Example 1 of the present disclosure are implemented.

[0123] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0124] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0125] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. The intelligent detection and identification method of tunnel defects based on visual perception is characterized by: The following steps are involved: Obtain a dataset of tunnel disease images and annotate the images; Based on the FOCS model, deformable convolutional layers are used to replace the third and fourth convolutions of the FOCS model backbone network and the first four convolutions of the two Head branches. The BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model. The DB-FOCS network model is trained using the annotated tunnel disease images to obtain a trained DB-FOCS network model; The tunnel disease image to be identified is input into the trained DB-FOCS network model, and the deformable convolution layer is used to extract features and generate deformed offsets. The deformed offsets are used to correct the extracted features and obtain multi-layer feature maps. The multi-layer feature maps are input into the BiFPN structure, and cross-layer feature fusion is performed to obtain cross-layer feature maps. The cross-layer feature maps are input into the Head branch, and the deformable convolution layer is used again to extract features to finally obtain the disease recognition results.

2. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 1, characterized in that: In the deformable convolutional layer: The traditional convolutional layer is used to extract the features of the tunnel disease image to be identified and obtain the initial feature map; Apply another convolution layer to the initial feature map to obtain the deformed offset, where the offset is 2N, which represents N 2-dimensional offsets, and N represents the number of convolution kernels; Based on the 2D offset, each pixel in the initial feature map is corrected to form N offset feature maps; Convolution is performed on the offset feature map to obtain a multi-layer feature map.

3. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 2, characterized in that: When training the deformable convolutional layer, the convolution kernel used to generate the output features and the convolution kernel used to generate the offset are learned synchronously, where the offset is learned by back propagation using an interpolation algorithm.

4. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 1, characterized in that: In the BiFPN structure: Upsample and downsample the input multi-layer feature maps respectively to make the feature maps of the same size; The information of each feature map is fused from top to bottom and bottom to top across multiple layers and scales; Add an extra edge between the input node and the output node to pass the information in the input feature map to the output; Each top-down and bottom-up bidirectional path is regarded as a feature network layer and repeated multiple times to obtain a cross-layer feature map.

5. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 4, characterized in that: An adjustable attention mechanism is used in the BiFPN structure to learn adjustable weights to control the flow of information in the feature map.

6. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 1, characterized in that: Input the cross-layer feature map into two Head branches: In the two head branches, the deformable convolutional layer is used to extract the features of the cross-layer feature map multiple times to obtain the first final feature map and the second final feature map; Using the classification submodule, a convolution operation is performed on the first final feature map to predict the boundary of the target for each feature point; The second final feature map is convolved by using the regression submodule and the center submodule, and the bounding box offset and size of the target and the center of the target are predicted for each feature point; The predicted target boundary, bounding box offset and size, and center are output as the defect recognition result.

7. The method for intelligent detection and identification of tunnel defects based on visual perception according to claim 1, characterized in that: The DB-FOCS network model uses the EIoU loss function: Among them, L IOU is the intersection-and-join loss; L dis Represents the distance loss between boxes; L asp represents the aspect ratio loss; ρ represents the weight factor; (b, b gt ) represents the position of the predicted box and the real box; (w, w gt ) represents the width of the predicted box and the real box, (h, h gt ) represents the height of the predicted box and the true box, c is the diagonal length of the smallest enclosing box covering the two boxes, C h and C w is the scaling factor for width and height loss.

8. The intelligent tunnel disease detection and identification system based on visual perception is characterized by: include: The data set acquisition module is configured to: acquire a tunnel disease image data set and annotate the images; The model building module is configured as follows: based on the FOCS model, the deformable convolutional layers are used to replace the third and fourth convolutions of the backbone network of the FOCS model, and the first four convolutions of the two Head branches, and the BiFPN structure is used to replace the FPN structure in the FOCS model to build the DB-FOCS network model; The training module is configured to: train the DB-FOCS network model using the annotated tunnel disease images to obtain a trained DB-FOCS network model; The disease recognition module is configured as follows: input the tunnel disease image to be identified into the trained DB-FOCS network model, use the deformable convolution layer to extract features and generate deformed offsets, use the deformed offsets to correct the extracted features and obtain a multi-layer feature map; input the multi-layer feature map into the BiFPN structure, perform cross-layer feature fusion, and obtain a cross-layer feature map; input the cross-layer feature map into the Head branch, use the deformable convolution layer to extract features again, and finally obtain the disease recognition result.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for intelligent detection and identification of tunnel defects based on visual perception as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for intelligent detection and identification of tunnel defects based on visual perception as described in any one of claims 1 to 7 are implemented.

Images