A method and system for detecting corrosion targets of enclosure structures

Abstract

The present invention provides a method and system for detecting corrosion targets of enclosure structures. By further enhancing the corrosion features extracted by a network model, the semantic information is enriched, which is beneficial for distinguishing background noise and defects, making the accuracy of corrosion detection higher and completing target detection in a shorter time, so that the enclosure structure corrosion target detection model can process large-scale image data and be used in real-time application scenarios.

Description

Technical Field

[0001] The invention belongs to the field of enclosure structure corrosion detection, and in particular relates to an enclosure structure corrosion target detection method and system. Background Art

[0002] The enclosure structure is an important part of the construction project. The corrosion detection of the existing enclosure structure is the key to ensure the safety and reliability of the structure. At present, the commonly used means is visual inspection. One of the simplest and most commonly used methods is to carefully observe the metal materials on the surface of the enclosure structure and look for any visible signs of corrosion, such as color changes, spots, oxides and corrosion pits. Visual inspection often requires professional staff to carry out under protective safety conditions. This method is not only time-consuming and laborious, but also accompanied by certain risks. The test results are also limited by the technical level of the inspectors. Summary of the invention

[0003] The main problem solved by the present invention is how to quickly and accurately detect corroded targets, and a method and system for detecting corroded targets of an enclosure structure are provided.

[0004] In order to solve the above technical problems, the technical solution adopted is:

[0005] A method for detecting corrosion targets of a protective structure comprises the following steps:

[0006] Step 1: Obtain the image of the enclosure structure;

[0007] Step 2: inputting the enclosure structure image into a trained enclosure structure corrosion target detection model to obtain a corrosion target;

[0008] The corrosion target detection model refers to a network model containing corrosion features for enhancing the corrosion features extracted from the enclosure structure image.

[0009] Furthermore, the method for enhancing the corrosion features extracted from the enclosure structure image refers to further extracting the corrosion features extracted from the network model.

[0010] Furthermore, a method for further extracting corrosion features from the corrosion features extracted by the network model is to use a corrosion feature enhancement module.

[0011] Furthermore, the feature enhancement method of the corrosion feature enhancement module is:

[0012] For the rust features extracted by the network model, the global average pooling operation GAP and the global maximum pooling operation GMP are used to compress the rust features of N channels into rust features of N vectors;

[0013] Add two N vectors channel by channel to form an N vector;

[0014] Divide the aggregated N vectors into K groups, and regard each group as a node;

[0015] Construct a fully connected graph with each node as a vertex and the relationships between nodes as edges;

[0016] Perform graphical reasoning on fully connected graphs;

[0017] The features of the graphic reasoning results are reconstructed, and the reconstructed features are used as corrosion features for corrosion target detection.

[0018] Furthermore, a method for performing graph reasoning on a fully connected graph is to use a graph neural network to perform graph reasoning on the fully connected graph.

[0019] Furthermore, the network model refers to an improved YoloV7 target detection network structure that uses an introduced feature pyramid network.

[0020] Furthermore, the rust feature enhancement module is added after the backbone feature extraction network of the YoloV7 target detection network structure. The features enhanced by the rust feature enhancement module are input into the feature pyramid, and the features extracted from the feature pyramid are input into the YOLO Head to obtain the rust target detection result and output it.

[0021] Furthermore, the method for training the corrosion target detection model is to use a training set for training, and the method for preparing the training set is:

[0022] The enclosure structure was selected as the test piece and a salt spray test was carried out in the laboratory to corrode the enclosure structure to different degrees;

[0023] Collect images of the enclosure structure at different corrosion levels;

[0024] The corrosion parts in the collected images are marked with features and a training set of enclosure structure corrosion is prepared.

[0025] The present invention also provides a system for detecting a target of corrosion of a protective structure, using various steps of a method for detecting a target of corrosion of a protective structure.

[0026] By adopting the above technical solution, the present invention has the following beneficial effects:

[0027] The present invention provides a method and system for detecting corrosion targets of enclosure structures. By further enhancing the corrosion features extracted by the network model, the semantic information is enriched, which is beneficial to distinguishing background noise and defects, making the accuracy of corrosion detection higher and completing target detection in a shorter time, so that the enclosure structure corrosion target detection model can process large-scale image data and be used in real-time application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Flowchart of the rust target detection method with the addition of the rust feature enhancement module;

[0029] Figure 2 This is a schematic diagram of the enhancement method of the corrosion enhancement module;

[0030] Figure 3 The process of making training sets and the training methods for network models. DETAILED DESCRIPTION

[0031] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. 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.

[0032] Figures 1 to 3 A specific embodiment of a method for detecting corrosion targets of a protective structure according to the present invention is shown, comprising the following steps:

[0033] Step 1: Obtain the building envelope image.

[0034] In this embodiment, professional photographing equipment is used to photograph the image of the enclosure structure.

[0035] Step 2: inputting the enclosure structure image into a trained enclosure structure corrosion target detection model to obtain a corrosion target;

[0036] The corrosion target detection model refers to a network model containing corrosion features for enhancing the corrosion features extracted from the enclosure structure image.

[0037] In this embodiment, Figure 1As shown, the network model refers to an improved YoloV7 target detection network structure using an introduced feature pyramid network. The rust feature enhancement module added to the improved YoloV7 target detection network structure is added after the main feature extraction network of the YoloV7 target detection network structure. The features enhanced by the rust feature enhancement module are input into the feature pyramid, and the features extracted from the feature pyramid are input into the YOLOHead to obtain the rust target detection results marked by the anchor frame and output.

[0038] The YoloV7 network structure uses a single-stage detection and anchor-free detection design idea, which only requires one forward propagation to complete target detection, so it has a faster detection speed. The YoloV7 network structure introduces the Feature Pyramid Network (FPN) to extract multi-scale features, which can more accurately detect rusted targets of different sizes. FPN combines feature maps from different levels and fuses information through upsampling and downsampling, so that the network can effectively detect rusted targets at different scales.

[0039] In this embodiment, the method for enhancing the corrosion features extracted from the image of the enclosure structure refers to further extracting the corrosion features extracted from the network model.

[0040] In this embodiment, the method for further extracting corrosion features from the corrosion features extracted by the network model is to use a corrosion feature enhancement module.

[0041] In this embodiment, Figure 2 As shown, the feature enhancement method of the corrosion feature enhancement module is:

[0042] 1) For the rust features extracted by the backbone feature network in the network model, the global average pooling operation GAP and the global maximum pooling operation GMP are used to compress the rust features of N channels into rust features of N vectors;

[0043] 2) Add the two N vectors channel by channel to form an N vector;

[0044] 3) Divide the aggregated N vectors into K groups, and regard each group as a node;

[0045] 4) Construct a fully connected graph with each node as a vertex and the relationships between nodes as edges;

[0046] 5) Performing graphical reasoning on the fully connected graph. In this embodiment, the method for performing graphical reasoning on the fully connected graph is to perform graphical reasoning on the fully connected graph using a graph neural network.

[0047] 6) Reconstruct features of the graphic reasoning results and use the reconstructed features as corrosion features to detect corrosion targets.

[0048] In this embodiment, the corrosion feature enhancement module (CFEM) performs further corrosion feature extraction on the result of feature extraction using the backbone feature extraction network as follows: Assuming that the features extracted by the backbone feature extraction network of the YoloV7 target detection network structure are Where N represents the number of channels, H represents the feature map height, and W represents the feature map width. First, in order to reduce the computational complexity of the rust feature enhancement module, feature C5 is converted to X , In the node representation stage, GAP and GMP are used to compress N channels into N vectors. , , and then aggregate their results to form a vector, , where n represents the nth channel in X, and X is represented by two N×1×1 vectors. Next, the aggregated N vectors are divided into K groups, each group is regarded as a node, and each node is represented by N / K channels. The combination of GAP and GMP captures the overall context and peak activation, providing an overall node representation for further node graph reasoning. Secondly, in the node graph reasoning stage, a fully connected graph is established, in which nodes represent vertices and their relationships are represented by edges in the graph. A graph neural network is used for graph reasoning. Represents the Laplacian matrix in the graph neural network, which is used to represent the information propagation between vertices in the graph and help capture the long-range dependencies between nodes. ,Represent2( X ( k )) is based on the characteristics of the node representation stage, X ( k ) indicates the k The features of nodes; the output of the graph neural network is , is the activation function, W g is a learnable weight matrix for node graph reasoning in graph neural networks. Finally, feature reconstruction is performed to combine X with G output Matrix multiplication results in Y, which represents the reconstructed features. .

[0049] The addition of the corrosion feature enhancement module CFEM enhances the global information modeling capability, enriches its semantic information, and is more conducive to distinguishing background noise from corrosion. It solves the problem of missed detection in traditional target detection tasks. The method of the present invention makes the accuracy of corrosion detection higher and can complete target detection in a faster time, which enables the enclosure structure corrosion target detection model to process large-scale image data and is suitable for real-time application scenarios.

[0050] In this embodiment, Figure 3 The method for training the rust target detection model is to use a training set for training, and the method for making the training set is:

[0051] 1) Select the corrugated steel sheet material commonly used in the enclosure structure as the test piece, and conduct a salt spray test in the laboratory to corrode the enclosure structure to different degrees. Due to the special nature of the enclosure structure, the data set can only be produced in the laboratory, so that rust samples in various stages of corrosion can be quickly obtained. Conduct a laboratory neutral salt spray test, the test is carried out in a salt spray test chamber, the test solution is NaCl solution, concentration 5%, pH 6.8; humidity>85%; the salt spray test machine setting parameters are: spray volume 1~2ml / h; air pressure 1.0+0.01kgf / cm 2 ; Pressure barrel temperature 47±1°C; Salt water barrel temperature 35±1°C; Sample placement angle 15~20°, spray mode continuous; Test specimens used 0.6mm thick Q235 corrugated steel plates commonly used in engineering, 45 specimens, the test lasted a total of 20 days (480h), and the accelerated corrosion corrugated steel plate specimens were collected after the test.

[0052] 2) Collect images of the enclosure structure at different corrosion levels. Use professional photography equipment to capture images of the test specimens; perform image preprocessing on the captured images, compile data enhancement algorithms using Python language, and perform data enhancement on the captured images. The data enhancement algorithms compiled during image preprocessing include but are not limited to using computer vision algorithms to perform rotation / reflection transformation, flip transformation, scaling transformation, translation transformation, etc. on the images to achieve the purpose of expanding the data set and improving the generalization ability and robustness of the model.

[0053] 3) Feature annotation of the corroded parts in the collected images to create a training set for the corrosion of the enclosure structure. In this embodiment, the feature annotation software labelimg is used to feature annotate the corrosion features in the data-enhanced images in the form of anchor frames to create a data set for corrosion target detection training.

[0054] The training set in the prepared rust target detection dataset is used to train the improved YoloV7 deep learning network with the rust feature enhancement module CFEM added. The test set is used to verify the accuracy of deep learning. Different groups of hyperparameters are set for multiple trainings, and the hyperparameters with the best accuracy are selected.

[0055] The present invention also provides a system for detecting a target of corrosion of a protective structure, using various steps of a method for detecting a target of corrosion of a protective structure.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting corrosion targets of enclosure structures, characterized in that: The following steps are involved: Step 1: Acquire the image of the enclosure structure; Step 2: inputting the enclosure structure image into a trained enclosure structure corrosion target detection model to obtain a corrosion target; The corrosion target detection model refers to a network model containing corrosion features for enhancing the corrosion features extracted from the enclosure structure image; The method for enhancing the rust features extracted from the image of the enclosure structure refers to further extracting the rust features extracted from the network model; The method of further extracting corrosion features from the corrosion features extracted by the network model is to use a corrosion feature enhancement module; The network model refers to an improved YoloV7 target detection network structure using an introduced feature pyramid network; The rust feature enhancement module is added after the main feature extraction network of the YoloV7 target detection network structure. The features enhanced by the rust feature enhancement module are input into the feature pyramid. The features extracted from the feature pyramid are input into the YOLO Head to obtain the rust target detection result and output it. The feature enhancement method of the corrosion feature enhancement module is: For the rust features extracted by the network model, the global average pooling operation GAP and the global maximum pooling operation GMP are used to compress the rust features of N channels into rust features of N vectors; Add two N vectors channel by channel to form a vector; , , respectively, represent the use of GAP and GMP to compress N channels into N vectors; H represents the feature map height, W represents the feature map width, and X refers to the vector after the feature C5 extracted by the backbone feature network is converted using 1×1 convolution, that is, , n means the nth channel is in X; Divide the aggregated vectors into K groups, and regard each group as a node; Construct a fully connected graph with each node as a vertex and the relationships between nodes as edges; Perform graphical reasoning on fully connected graphs; The features of the graphic reasoning results are reconstructed, and the reconstructed features are used as corrosion features for corrosion target detection.

2. The target detection method according to claim 1, characterized in that: The method of performing graph reasoning on fully connected graphs is to use graph neural networks to perform graph reasoning on fully connected graphs.

3. The target detection method according to claim 2, characterized in that: The method for training the rust target detection model is to use a training set for training, and the method for preparing the training set is: The enclosure structure was selected as the test piece and a salt spray test was carried out in the laboratory to corrode the enclosure structure to different degrees; Collect images of the enclosure structure at different corrosion levels; The corrosion parts in the collected images are marked with features and a training set of enclosure structure corrosion is prepared.

4. A corrosion target detection system for enclosure structures, characterized in that: Use each step of a method for detecting corrosion targets of a protective structure as described in any one of claims 1 to 3.

Images