Structure multi-disease identification method based on lightweight full convolution residual network

Abstract

The invention relates to a structure multi-disease identification method based on a lightweight full convolution residual network, and the method comprises the following steps: obtaining an image of a single disease, setting the size of the image, marking each image sample, and generating a single disease data set; preprocessing images in the single disease data set, generating a compound disease sample image in which multiple diseases coexist through cross aliasing, and constructing a multi-disease data set; building a full convolution residual network model, setting hyper-parameters, and training the network by adopting an Adam optimization algorithm and a classification cross entropy loss function; and carrying out structure disease identification by using the trained full convolution residual network, thereby realizing simultaneous, rapid and accurate identification of multiple diseases. Compared with the prior art, various diseases can be efficiently identified under the condition that various diseases coexist, the limitation that a traditional method can only identify a single disease is overcome, and the method has remarkable practical value and a good application effect.

Description

Technical Field

[0001] The present invention relates to the field of structural health monitoring and computer deep learning technology, and in particular to a structural multi-disease recognition method based on a lightweight fully convolutional residual network. Background Art

[0002] In large-scale engineering structures such as highways, bridges, high-rise buildings, airport runways, tunnels and other buildings, various surface defects such as damage, cracking, corrosion, and shedding often occur on their surfaces due to climate, environment, collision and use of vehicles. The presence of surface defects will lead to a decrease in the strength, stiffness, durability and other properties of the structure, thereby increasing the risk of accidents. At the same time, maintaining and repairing surface defects is also a huge workload and cost, so it is very important to detect and deal with surface defects as early as possible. Structural surface defect identification is to identify and monitor the occurrence and development of surface defects, obtain effective information in a timely manner, and repair them in a timely manner.

[0003] Traditional structural defect identification methods mainly rely on visual inspection, that is, identifying defects by observing the surface of the structure with the naked eye. Although this method is direct, it is inefficient and requires a lot of time and human resources. With the advancement of scientific and technological advances such as sensor technology, electronic measurement and image processing, structural defect detection technology has been significantly improved, especially in the use of neural network technology to process images, which can significantly improve the recognition accuracy of structural defect images.

[0004] However, existing technologies have some limitations in the identification of structural defects. First, these technologies are mainly used to identify single defects. For multiple defects that occur simultaneously on a structure, i.e., structural multi-defects, there is a lack of in-depth research and effective identification methods. Secondly, there are relatively few samples of structural multi-defects, especially compound defects, i.e., multiple defects exist simultaneously in the same structural area (for example, the concrete surface has both honeycomb defects and cracks), which makes traditional methods obviously insufficient in identifying compound defects. In addition, due to the seldom coexistence of multiple structural defects, relevant data sets are difficult to obtain. Most of the current research and applications focus on single-defect data sets, which limits the development and application of multi-defect identification technology. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a structural multi-disease recognition method based on a lightweight fully convolutional residual network, which can efficiently identify various diseases when multiple diseases coexist, overcomes the limitation that traditional methods can only identify a single disease, and has significant practical value and good application effect.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] The present invention provides a method for identifying multiple structural defects based on a lightweight fully convolutional residual network, comprising the following steps:

[0008] S1: Construction and labeling of single disease dataset: Obtain single disease images and set the image size, label each image sample to form a single disease dataset;

[0009] S2: Construction and labeling of multi-disease dataset: By preprocessing the single disease images in the single disease dataset, composite disease sample images with multiple diseases coexisting are generated through cross-aliasing to form a multi-disease dataset;

[0010] S3: Build a fully convolutional residual network model, including convolutional layers, pooling layers, classification layers, full quality attention blocks, and point-by-point feature selection blocks. Based on the multi-disease dataset, set hyperparameters, and use the Adam optimization algorithm and classification cross entropy loss function to train the network.

[0011] S4: Multiple structural defects are identified based on the trained fully convolutional residual network, achieving rapid and accurate identification of multiple structural defects at the same time.

[0012] Furthermore, in S1, the specific process includes:

[0013] Get a single disease image and set the image sample size;

[0014] Each image sample is labeled, including: the complete structure surface is marked as 0, the block loss of the structure surface is marked as 1, the honeycomb on the structure surface is marked as 2, the structure crack is marked as 3, etc., a total of N types, and the result corresponding to the input X is Y, where each element y_i of Y corresponds to the sample x_i in X, and the value range of y_i is 0 to N-1, and m is the number of samples.

[0015] Furthermore, in S2, the preprocessing includes: performing at least one of scaling, stretching, flipping, rotating, cropping, and adding noise to the single defect image.

[0016] Furthermore, in S2, in the cross-aliasing process, a random linear mixing method is used to obtain a multi-disease network data set and its tags Specifically include:

[0017] Randomly select u single disease images from the m single disease images in the single disease dataset for construction;

[0018] Linearly mix u single disease images and combine u single disease image labels to obtain new labels:

[0019] Repeat the above process m times to get the final and

[0020] Furthermore, in S3, the process of building a fully convolutional residual network includes the construction of convolutional blocks, specifically including:

[0021] Set the input multi-disease image data format to (C0, H0, W0), and define the multi-disease image output data format to (C, H, W);

[0022] Performing two-dimensional convolution, batch normalization, ReLU activation function, two-dimensional convolution and batch normalization operations on a portion of the input multi-disease image data;

[0023] Only two-dimensional convolution and batch normalization operations are performed on another part of the input multi-disease image data;

[0024] The two results are superimposed and the same convolution and batch normalization operations are performed again to obtain the convolution block output.

[0025] Furthermore, in S3, the process of building a fully convolutional residual network includes the construction of a point-by-point feature selection block, specifically including:

[0026] Analyze the input feature map, covering the three dimensions of number of channels, height and width of multi-disease image features;

[0027] Divide the feature map into three subgroups based on channels, retain the features in the highest and middle subgroups, and discard the features in the lowest subgroup;

[0028] The remaining two subgroups are averaged point by point and compressed into an independent subgroup, keeping the spatial dimension of the feature map unchanged.

[0029] Furthermore, in S3, the process of building a fully convolutional residual network includes the construction of a full-quality attention block, specifically including:

[0030] In the channel attention path, the channel attention map is generated through a skip layer and a bottleneck structure, where the bottleneck structure contains two dense layers activated by ReLU and sigmoid activation functions respectively;

[0031] In the spatial attention path, spatial descriptors along the channel axis are computed, and spatial attention maps are generated using convolutional layers and transposed convolutional layers;

[0032] The channel and spatial attention maps are element-wise multiplied with the input feature map to weight the importance of the features.

[0033] Furthermore, in S3, the process of building a fully convolutional residual network also includes the construction of the FCResNet neural network framework, which specifically includes:

[0034] Perform 2D convolution, batch normalization, ReLU activation function, and max pooling operations on the training network input;

[0035] Construct a network structure containing 4 convolutional blocks;

[0036] After the four convolution blocks, a 2D adaptive average pooling layer, a 2D convolution, a batch normalization, and a Softmax classification layer are added;

[0037] Define the Softmax classification layer to calculate the probability distribution.

[0038] Furthermore, in S3, the expression of the maximum pooling operation is:

[0039]

[0040] in is the jth channel of the lth layer; P i l is the i-th channel of the pooling operator in the l-th layer, I is the channel volume of layers l-1 and l, and the in-plane size calculation expression of the output feature map is as follows:

[0041] D out =(D in -P pool ) / S pool +1

[0042] Where P pool is the side length of the pooling operation, S pool is the stride length, D in is the feature map input of the pooling layer, D out It is the feature map output of the pooling layer, and the maximum pooling is the maximum value of the calculation area;

[0043] The four convolution blocks are followed by a 2D adaptive average pooling layer, a 2D convolution, a batch normalization, and a Softmax classification layer;

[0044] The Softmax classification layer is defined as:

[0045]

[0046] in, and 0≤softmax(x) i ≤1, softmax(x) i is a probability, k represents the number of categories; xi represents the input.

[0047] Furthermore, in S3, the classification cross entropy loss function is defined as follows:

[0048]

[0049] Among them, the vector According to the label data, make its y(i) elements are 1, and the rest are 0; is the predicted probability distribution; m represents the number of samples in the training data set, and k represents the number of categories; and They are y (i) and The elements in are either 0 or 1; Θ represents the model parameters.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1) The present invention successfully simulates and constructs compound disease samples through innovative cross-aliasing technology. This technology superimposes multiple single disease images through image processing to form simulated compound disease images. The present invention effectively solves the problem of scarce compound disease samples and provides sufficient training data for the deep learning model, thereby improving the accuracy and generalization ability of the model in identifying compound diseases.

[0052] 2) The FCResNet network of the present invention can identify and classify a variety of structural diseases including single diseases and compound diseases. Through the setting of fully convolutional residual networks with different numbers of layers, the present invention not only improves the accuracy of traditional residual networks, but also optimizes the identification of compound diseases, so that the model can accurately identify multiple diseases occurring simultaneously on the structure in practical applications, significantly improving the efficiency and accuracy of structural disease identification.

[0053] 3) The method proposed in the present invention of constructing a multi-disease dataset using image stitching technology provides a new data source for the training of deep learning models. The present invention enhances the model's ability to recognize complex disease scenarios by simulating actual possible compound disease situations.

[0054] 4) The present invention also has the advantages of being not susceptible to interference and being highly efficient, which reduces a lot of manpower and improves work efficiency. At the same time, the method can realize the simultaneous identification of multiple structural defects. In addition, the method proposes a method of constructing a multi-disease data set using image stitching technology, which solves the problem that the multi-disease data set is small and difficult to train. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a framework diagram of the structural defect recognition method based on residual fully convolutional neural network (FCResNet);

[0056] Figure 2 It is the convolutional block architecture diagram;

[0057] Figure 3 It is a point-by-point feature selection block architecture diagram;

[0058] Figure 4 This is the full quality attention block architecture diagram;

[0059] Figure 5 It is the residual fully convolutional neural network architecture diagram. DETAILED DESCRIPTION

[0060] On the whole, the present invention discloses a method for structural multi-defect recognition based on residual full convolutional neural network (FCResNet). The method introduces a lightweight attention network with multi-scale feature fusion to solve the challenge of high recognition computational complexity, and introduces full-quality attention blocks and point-by-point feature selection blocks to identify multi-defect features. The method first labels the directly collected structural surface image data set, then expands and enhances the multi-defect samples through image cross-aliasing technology, reconstructs the network data set and labels it, and then divides the new image data set into a training set and a test set, and then builds a FCResNet network for structural multi-defect recognition. The training and test sets of the method are used for network training and verification of the accuracy of the method. The method is based on a supervised deep learning network, including a convolution layer, a pooling layer, a classification layer, a full-quality attention block and a point-by-point feature selection block. The artificially marked abnormal classification is compared with the predicted output of the test set data of the FCResNet network to judge the excellence of the FCReNet network, and finally the model with good training accuracy is saved for predicting data and completing classification. This method can realize the rapid and accurate identification of multiple structural diseases at the same time, and overcomes the limitation of existing methods that can only identify a single disease, and has achieved good results.

[0061] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms and other features not clearly described in this technical solution are all considered to be common technical features disclosed in the prior art.

[0062] Example 1

[0063] In this embodiment, the structural defect recognition method based on the fully convolutional residual network has the following process: Figure 1 As shown, the following steps are included:

[0064] 1) Dataset construction;

[0065] 2) Dataset labeling;

[0066] 3) Build the FCResNet network and set hyperparameters for training;

[0067] 4) Structural defect recognition based on FCResNet network.

[0068] The process of step 1) is as follows:

[0069] Get a single defect image, set the image size, and assume that the data format of input X is (m,c,h,w), where m represents the number of samples, that is, the number of input images, c represents the number of channels, h represents the image height, and w represents the image width. Assuming that the image length and width are the same, both h, the data format of X is (m,c,h,h).

[0070] Each single defect image sample is labeled. Assume that there are N types in total, where the structural surface is intact (0), the structural surface is broken (1), the structural surface has honeycombs (2), the structural cracks (3), etc. The result corresponding to the input X is Y:

[0071] Y=[y1,y2,…y i ,…y m ] Τ (1)

[0072] Among them, y i ∈{0, 1,…, N-1}, m is the number of samples.

[0073] The process of step 2) is as follows:

[0074] The original single disease image dataset obtained in step 1) is subjected to random linear mixing to obtain a multi-disease network dataset. and its tags The specific process is:

[0075] ① Randomly select u images from m single-disease images for construction. For the convenience of calculation, u is taken as a square number;

[0076] ② Linearly mix u single-disease images. The following takes u=9 as an example:

[0077]

[0078] ③ Combine u single disease image labels to obtain a new label:

[0079] y new =[y1,y2,…,y9] (3)

[0080] ④ Repeat the above process m times to obtain and The data format is (m,c,3h,3h), and its corresponding label

[0081]

[0082] Among them, y i_new The calculation method of is shown in process ③.

[0083] The process of step 3) is as follows:

[0084] 3.1 Convolutional Block Construction

[0085] See also Figure 2 , construct a convolution module, assuming that the input data format of the multi-disease image is (C0, H0, W0), and the output data format of the multi-disease image is (C, H, W), where the input and output formats can be the same or different. One part of the input is subjected to two-dimensional convolution, batch normalization, ReLU activation function, two-dimensional convolution and batch normalization operations, and the other part is subjected to two-dimensional convolution and batch normalization operations. The results of the two parts are superimposed, and the above steps are performed again to obtain the convolution block output.

[0086] The convolution operation is as follows:

[0087]

[0088] in is the jth channel of the lth layer, is the ith channel of filter j in layer l, is the bias of filter j in layer l, and I and J are the channel sizes of layers l-1 and l.

[0089] When performing a convolution operation on an image, the filter performs an element-wise multiplication with the receptive field of the input, sums the products in each channel, and then advances by the specified stride, and so on. The in-plane size of the output feature map is calculated by the following formula:

[0090] D out =(D in -C) / S conv +1 (6)

[0091] Where D in is the input layer x l-1 Here, we assume that the input layer is a layer with size D in each channel. in ×D in C is the side length of the filter. Here we assume that each filter is a square of size C×C. out is the output layer x l The side length of S conv is the stride, assuming here that the stride along the height and width directions is S conv .

[0092] The batch normalization operation is to standardize the input data, that is, to change the sample mean to 0 and the standard deviation to 1.

[0093] The ReLU activation function is expressed as:

[0094]

[0095] 3.2 Point-wise feature selection block construction

[0096] Existing methods generally adopt the direct fusion of multi-scale features and assign uniform weights to these features. However, this direct fusion feature extraction method often contains redundant information, which produces a blurring effect on the presentation of key details. Especially in complex real-world environments, the presence of a large number of redundant details in the background area further exacerbates this problem. This redundant information not only increases the difficulty of structural multi-hazard identification, but may also significantly reduce the performance of the identification system.

[0097] The present invention introduces a point-by-point feature selection block, which reduces redundant information by eliminating weak features at each scale one by one before fusion. This feature selection module effectively reduces the number of parameters, making the system architecture more efficient and compact. In addition, this method can extract more informative features from relatively few feature maps, significantly improving the generalization ability of the FCResNet model.

[0098] The point-wise feature selection block starts with the analysis of the input feature map, covering the three dimensions of the number of channels, height, and width of the feature. The feature map is then divided into three subgroups based on the channel, and the features in the highest and middle subgroups are retained, while the lowest feature in all subgroups is discarded. The remaining two subgroups are then averaged point-wise and compressed into a single subgroup. In this process, the spatial dimension of the feature map remains unchanged, thereby optimizing the feature representation and improving the performance of the system.

[0099] 3.3 Full-quality attention block construction

[0100] See also Figure 4 ,The full quality attention block applies the attention mechanism by ,dynamically adjusting the importance of feature maps in both channel and ,spatial dimensions, which significantly improves the accuracy of multi-hazard feature recognition.,The input and output processing flow of this module consists of ,two parallel paths.

[0101] In the upper path, i.e., the channel attention path, the channel attention map is generated through the skip layer and the bottleneck structure. The skip layer first generates the channel description information of the feature map, and then inputs it into the bottleneck structure. The bottleneck structure contains two dense layers, which are activated by ReLU and sigmoid activation functions respectively. The first dense layer reduces the channel dimension to one-fourth of the original input, while the second dense layer restores it to the original dimension, similar to the self-attention mechanism, accurately focusing on important channel information.

[0102] The lower path, the spatial attention path, uses the same process to generate the spatial attention map. First, the spatial descriptor along the channel axis is calculated, which takes the average of the input feature map through a point-by-point jump operation. Next, the spatial dimension is gradually reduced using two convolutional layers with a stride of 2, and then restored to the original spatial size through two transposed convolutional layers. The final stage multiplies the channel and spatial attention maps with the input feature map element-wise to achieve feature importance weighting. Through this dual-path, hierarchical attention mechanism, the full-quality attention block enhances the model's ability to capture important features, enabling it to excel in the complex task of identifying multi-disease features.

[0103] 3.4FCresNet neural network framework construction

[0104] See also Figure 5 , construct the model structure of FCResNet neural network, train the network input to first perform two-dimensional convolution, batch normalization, ReLU activation function and maximum pooling operations, and then train the network structure to include four convolution modules, each of which is shown in 3.1. The expression of the pooling operation is:

[0105]

[0106] in is the jth channel of the lth layer; P i l is the i-th channel of the pooling operator in layer l; I is the number of channels in layers l-1 and l. The in-plane size of the output feature map is calculated as follows:

[0107] D out =(D in -P pool ) / S pool +1 (9)

[0108] Where P pool is the side length of the pooling operation (assuming the operator is P pool ×P pool square), S pool is the stride (both height and width are S pool ), D in is the feature map input of the pooling layer, D out It is the feature map output of the pooling layer. Max pooling is the maximum value of the calculation area.

[0109] The four convolution blocks are followed by a 2D adaptive average pooling layer, a 2D convolution, a batch normalization, and a Softmax classification layer; the Softmax classification layer is defined as (10):

[0110]

[0111] in, and 0≤softmax(x) i ≤1, softmax(x) i is a probability, k represents the number of categories; x i Represents input;

[0112] The Adam optimization algorithm is used for network training; the loss function of network training adopts classification cross entropy, which is defined as follows:

[0113]

[0114] Among them, the vector According to the label data, make its y (i) elements are 1, and the rest are 0; is the predicted probability distribution; m represents the number of samples in the training data set, and k represents the number of categories; and They are y (i) and The elements in are either 0 or 1; Θ represents the model parameters;

[0115] 3.5 Save the trained FCResNet model;

[0116] Perform a model quality test on the test set and save the model with excellent results.

[0117] The process of step 4) is as follows:

[0118] The multi-disease dataset and labels obtained in step 2) are brought into the FCResNet model constructed in step 3) for prediction. The specific process is as follows:

[0119] ①Extract each sample x from the data set i and the corresponding label y i ;

[0120] ②x i The image sample has a width and height of 3h, which are much larger than the model input width and height h. The sample is cut into pieces by h to obtain 9 prediction samples. The prediction input format is (9, 3, h, h):

[0121] ③ Input the prediction sample into the model to obtain the prediction result in To predict the results,

[0122] ④According to the prediction results, the identification results of multiple structural defects are obtained.

[0123] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for structural multi-defect recognition based on lightweight fully convolutional residual network, characterized in that: The following steps are involved: S1: Construction and labeling of single disease dataset: Obtain single disease images and set the image size, label each image sample to form a single disease dataset; S2: Construction and labeling of multi-disease dataset: By preprocessing the single disease images in the single disease dataset, composite disease sample images with multiple diseases coexisting are generated through cross-aliasing to form a multi-disease dataset; S3: Build a fully convolutional residual network model, including convolutional layers, pooling layers, classification layers, full quality attention blocks, and point-by-point feature selection blocks. Based on the multi-disease dataset, set hyperparameters, and use the Adam optimization algorithm and classification cross entropy loss function to train the network. S4: Multiple structural defects are identified based on the trained fully convolutional residual network, achieving rapid and accurate identification of multiple structural defects at the same time.

2. According to claim 1, a method for identifying multiple structural defects based on a lightweight fully convolutional residual network is characterized in that: In S1, the specific process includes: Get a single disease image and set the image sample size; Each single defect image sample is labeled, including: the complete structure surface is marked as 0, the block loss of the structure surface is marked as 1, the honeycomb on the structure surface is marked as 2, the structure crack is marked as 3, etc., a total of N types, and the result corresponding to the input X is Y, where each element y_i of Y corresponds to the sample x_i in X, and the value range of y_i is 0 to N-1, and m is the number of samples.

3. According to claim 1, a method for identifying multiple structural defects based on a lightweight fully convolutional residual network is characterized in that: In S2, the preprocessing includes: performing at least one of scaling, stretching, flipping, rotating, cropping, and adding noise to the single defect image.

4. According to claim 1, a method for identifying multiple structural defects based on a lightweight fully convolutional residual network is characterized in that: In S2, during the cross-aliasing process, a random linear mixing method is used to obtain a multi-disease network dataset and its tags Specifically include: Randomly select u single disease images from the m single disease images in the single disease dataset for construction; Linearly mix u single disease images and combine u single disease image labels to obtain new labels: Repeat the above process m times to get the final and 5. According to claim 1, a method for identifying multiple structural defects based on a lightweight fully convolutional residual network is characterized in that: In S3, the process of building a fully convolutional residual network includes the construction of convolutional blocks, specifically: Set the input multi-disease image data format to (C0, H0, W0), and define the output multi-disease image data format to (C, H, W); Performing two-dimensional convolution, batch normalization, ReLU activation function, two-dimensional convolution and batch normalization operations on a portion of the input multi-disease image data; Only two-dimensional convolution and batch normalization operations are performed on another part of the input multi-disease image data; The two results are superimposed and the same convolution and batch normalization operations are performed again to obtain the convolution block output.

6. According to claim 1, a method for identifying multiple structural defects based on a lightweight fully convolutional residual network is characterized in that: In S3, the process of building a fully convolutional residual network includes the construction of point-by-point feature selection blocks, specifically including: Analyze the input multi-disease image feature map, covering the three dimensions of the number of channels, height and width of the multi-disease image features; The multi-disease feature map was divided into three subgroups based on the channel, the features in the highest and middle subgroups were retained, and the features in the lowest subgroup were discarded; The remaining two subgroups are averaged point by point and compressed into an independent subgroup, keeping the spatial dimension of the feature map unchanged.

7. The method for structural multi-defect recognition based on lightweight fully convolutional residual network according to claim 1, characterized in that: In S3, the process of building a fully convolutional residual network includes the construction of a full-quality attention block, specifically including: In the channel attention path, the channel attention map is generated through a skip layer and a bottleneck structure, where the bottleneck structure contains two dense layers activated by ReLU and sigmoid activation functions respectively; In the spatial attention path, spatial descriptors along the channel axis are computed, and spatial attention maps are generated using convolutional layers and transposed convolutional layers; The channel and spatial attention maps are element-wise multiplied with the input feature map to weight the importance of the features.

8. The method for structural multi-defect recognition based on lightweight fully convolutional residual network according to claim 1, characterized in that: In S3, the process of building a fully convolutional residual network also includes the construction of the FCResNet neural network framework, which includes: Perform 2D convolution, batch normalization, ReLU activation function, and max pooling operations on the training network input; Construct a network structure containing 4 convolutional blocks; After the four convolution blocks, a 2D adaptive average pooling layer, a 2D convolution, a batch normalization, and a Softmax classification layer are added; Define the Softmax classification layer to calculate the probability distribution.

9. The method for structural multi-defect recognition based on lightweight fully convolutional residual network according to claim 8, characterized in that: In S3, the expression of the maximum pooling operation is: in is the jth channel of the lth layer; P i l is the i-th channel of the pooling operator in the l-th layer, I is the channel volume of layers l-1 and l, and the in-plane size calculation expression of the output feature map is as follows: D out =(D in -P pool ) / S pool +1 Where P pool is the side length of the pooling operation, S pool is the stride length, D in is the feature map input of the pooling layer, D out It is the feature map output of the pooling layer, and the maximum pooling is the maximum value of the calculation area; The four convolution blocks are followed by a 2D adaptive average pooling layer, a 2D convolution, a batch normalization, and a Softmax classification layer; The Softmax classification layer is defined as: in, and 0≤softmax(x) i ≤1, softmax(x) i is a probability, k represents the number of categories; xi represents the input.

10. The method for structural multi-defect recognition based on lightweight fully convolutional residual network according to claim 1, characterized in that: In S3, the classification cross entropy loss function is defined as follows: Among them, the vector According to the label data, make its y (i) elements are 1, and the rest are 0; is the predicted probability distribution; m represents the number of samples in the training data set, and k represents the number of categories; and They are y (i) and The elements in are either 0 or 1; Θ represents the model parameters.

Images