A low-temperature flexible detection method for waterproofing membrane based on image recognition

Abstract

The present application relates to a kind of low-temperature flexible detection methods of waterproofing membrane based on image recognition, belong to image intelligent identification technical field, solve the problem that existing waterproofing membrane cannot be accurately and quickly identified after low-temperature flexible detection Crack of membrane.The method comprises: obtaining the image of multiple waterproofing membrane samples after low-temperature flexible test;Each image is respectively input into the trained waterproofing membrane crack detection model to detect the crack of bending part, and the crack information of each sample bending part is obtained;Wherein, the waterproofing membrane crack detection model includes bending detection submodule and crack detection submodule;The bending detection submodule is used to detect the bending part in sample;The crack detection submodule is used to detect the crack information of the bending part;When the number of samples with cracks exceeds the preset value, it is judged that the low-temperature flexible test is unqualified.The low-temperature flexible detection of waterproofing membrane is realized to the crack in the fast and accurate identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image intelligent recognition, and particularly relates to a low-temperature flexible detection method for waterproofing membrane based on image recognition. BACKGROUND

[0002] With the development of modern architecture, waterproofing membranes are increasingly widely used in the field of building waterproofing such as roofs and basements. However, in some cold regions, the low-temperature environment can have a certain negative impact on the use of waterproofing membranes. In a low-temperature environment, the flexibility of waterproofing membranes can be poor, and cracking and breaking can easily occur, thereby affecting the service life and performance of the waterproofing membranes. In order to protect the service life and performance of the waterproofing membranes, low-temperature flexibility tests must be performed. The low-temperature flexibility test mainly simulates the actual situation in a low-temperature environment, and performs bending, stretching and other tests on the waterproofing membranes to evaluate the flexibility and toughness of the waterproofing membranes in a low-temperature environment.

[0003] After the existing low-temperature flexibility detection equipment for waterproofing membranes completes the standard detection, the detection personnel takes out the sample of the waterproofing membrane, and uses the naked eye to determine whether the membrane is cracked, so as to determine whether the test batch is qualified. Since the low-temperature flexibility of the waterproofing membrane is an important performance indicator of the waterproofing membrane, the demand for laboratory and site detection is large in engineering construction, and the judgment of small cracks is often biased due to the limitation of the light test conditions and the influence of human fatigue factors. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a low-temperature flexibility detection method for waterproofing membranes based on image recognition, to solve the problem that the existing waterproofing membranes cannot accurately and quickly identify the cracks of the membranes after low-temperature flexibility detection.

[0005] The main purpose of the present application is achieved by the following technical solutions:

[0006] The present application provides a low-temperature flexibility detection method for waterproofing membranes based on image recognition, comprising the following steps:

[0007] Obtaining images of a plurality of waterproofing membrane samples subjected to low-temperature flexibility tests;

[0008] inputting each image into a trained waterproofing membrane crack detection model to detect cracks in the bending part, and obtaining crack information of the bending part of each sample; wherein the waterproofing membrane crack detection model comprises a bending detection submodule and a crack detection submodule; the bending detection submodule is used to detect the bending part in the sample; the crack detection submodule is used to detect crack information of the bending part;

[0009] When the number of samples with cracks exceeds a preset value, it is determined that the low-temperature flexibility test is unqualified.

[0010] Further, the crack detection sub-module comprises a high-frequency feature extraction module, a spatial domain feature extraction module and a frequency domain and spatial domain feature mixing module.

[0011] The images of the bending parts are respectively input into the high-frequency feature extraction module and the spatial domain feature extraction module of the crack detection sub-module, to obtain high-frequency features and RGB features of different scales of the images of the bending parts.

[0012] After the high-frequency features and the RGB features of the same scale are fused in the frequency domain and spatial domain feature mixing module of the crack detection sub-module, the crack information in the images of the bending parts is output by the crack detection head.

[0013] Further, the high-frequency features and the RGB features are fused in the frequency domain and spatial domain feature mixing module of the crack detection sub-module, comprising:

[0014] The high-frequency features and the RGB features are multiplied by matrices to obtain mixed frequency domain features.

[0015] The mixed frequency domain features are respectively multiplied by the high-frequency features and the RGB features to obtain high-frequency enhanced features and RGB enhanced features.

[0016] The high-frequency enhanced features and the RGB features are added by matrices to obtain high-frequency enhanced-RGB fusion features.

[0017] The RGB enhanced features and the high-frequency features are added by matrices to obtain RGB enhanced-high-frequency fusion features.

[0018] The high-frequency enhanced-RGB fusion features and the RGB enhanced-high-frequency fusion features are channel fused by matrices to obtain fusion features.

[0019] Further, the images of the bending parts in the images are sequentially input into the backbone network layer and the multi-scale feature fusion layer of the spatial domain feature extraction module to obtain RGB features of different scales; wherein the multi-scale feature fusion layer comprises three target detection layers of different scales.

[0020] Further, in the high-frequency feature extraction module, the images of the bending parts are converted from the spatial domain to the frequency domain by discrete cosine transformation, and the high-frequency region weight of the image is increased by a filter to obtain a high-frequency enhanced image; the high-frequency enhanced image is restored to the spatial domain by inverse discrete cosine transformation, and then passes through three convolution layers to obtain high-frequency features of three different scales.

[0021] Further, the waterproof coiled material crack detection model is obtained by training the following method:

[0022] respectively, wherein the bending detection training dataset comprises sample images and labeled bending detection boxes; and the crack detection training dataset comprises images in the bending detection boxes and corresponding crack information;

[0023] loading the bending detection training dataset to train the bending detection submodule to obtain a trained bending detection submodule;

[0024] pre-training a backbone network of a spatial feature extraction module of the crack detection submodule based on contrastive learning to obtain a pre-trained backbone network;

[0025] loading the crack detection training dataset to train the crack detection submodule containing the pre-trained backbone network to obtain a trained crack detection submodule;

[0026] based on the trained bending detection submodule and the trained crack detection submodule, obtaining a trained waterproofing membrane crack detection model.

[0027] Further, the pre-training of the backbone network of the spatial feature extraction module of the crack detection submodule based on contrastive learning to obtain a crack detection basic backbone network comprises:

[0028] a plurality of training data groups are constructed, each training data group comprising a plurality of training sample pairs; wherein the training sample pair comprises a positive sample and a negative sample; the positive sample in each training data group is a picture of the same crack region, and the negative sample is a different picture without a crack region;

[0029] for each training data group, loading the training sample pair, training the backbone network of the spatial feature extraction module of the crack detection submodule using a contrastive loss function, updating the backbone network parameters of the spatial feature extraction module using gradient backpropagation, and saving the backbone network parameters of the spatial feature extraction module after training.

[0030] Further, the negative sample in each training data group comprises a plurality of crack-free region pictures of the same size as the positive sample and two pictures of the same size as the positive sample extracted on both sides of the long side of the positive sample.

[0031] Further, the loading of the bending detection training dataset to train the bending detection submodule to obtain a trained bending detection submodule comprises: loading the bending detection training dataset, training the bending detection submodule using a combination of a bounding box intersection over union loss function, a classification loss function and a distribution focal loss function, updating the parameters of the bending detection submodule using a stochastic gradient descent method for backpropagation, and obtaining a trained bending detection submodule;

[0032] The loading of the crack detection training data set trains the crack detection sub-module containing the pre-trained backbone network to obtain a trained crack detection sub-module, comprising: loading the crack detection training data set, using the combination of the boundary box small target detection loss function, the classification loss function and the distribution focal loss function, training the crack detection sub-module, using the random gradient descent method to update the parameters of the crack detection sub-module by back propagation, and obtaining the trained crack detection sub-module.

[0033] Further, the structure of the bending detection sub-module is a YOLOv9 model structure.

[0034] Compared with the prior art, the present application can at least achieve one of the following beneficial effects:

[0035] 1. For waterproofing membranes, cracks usually occur at the bending part of the waterproofing membrane, and the waterproofing membrane crack detection model of the present application first detects the bending part of the waterproofing membrane using the bending detection sub-module, and then further judges whether there is a crack using the crack detection sub-module, which is more targeted than directly detecting cracks on the waterproofing membrane image, and can improve the detection accuracy of these areas.

[0036] 2. The waterproofing membrane crack detection model of the present application uses the bending detection sub-module to detect the bending part of the waterproofing membrane, which can reduce the detection in areas where no cracks occur, thereby reducing false positives due to random textures or material non-uniformity, and reducing the image range that needs to be detected at high resolution, thereby saving computing resources and time.

[0037] 3. Since for industrial scenes, the samples of waterproofing membranes are usually imbalanced, i.e. the number of normal waterproofing membrane samples is much larger than the number of waterproofing membrane crack samples, therefore, when training the crack detection sub-module of the waterproofing membrane crack detection model, the present application performs pre-training of the backbone network layer based on contrastive learning, so that the backbone network layer of the crack detection sub-module learns the intra-class consistency and inter-class difference feature representation of the waterproofing membrane crack, improves the detection effect of the waterproofing membrane crack detection model, and reduces model overfitting.

[0038] 4. When constructing the training sample pair for pre-training of the backbone network layer of the crack detection sub-module of the waterproofing membrane crack detection model based on contrastive learning, the negative sample includes two images with the same size as the positive sample extracted on both sides of the long side of the positive sample in the same training data set, by selecting an area similar to the crack area but actually containing no crack as a negative sample, the model can learn the subtle differences between the crack and the normal area, enhance the difference of the features, and more accurately identify the crack area.

[0039] 5、The crack detection sub-module of the waterproof roll crack detection model in the application extracts high-frequency feature information in the roll image that is difficult to detect in the RGB space, can better identify the existence of cracks, enhance the feature performance of small cracks, and thus improve the detection capability of the model for small cracks.

[0040] The various technical solutions described above can be combined with each other to achieve more preferred combination solutions. Other features and advantages of the application will be described in the subsequent description, and some advantages will become apparent from the description or will be understood by implementing the application. The purposes and other advantages of the application can be achieved and obtained from the contents specifically pointed out in the description and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings are included to provide a further understanding of the embodiments, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the principles of the application. In the drawings:

[0042] Figure 1 A flowchart of a low-temperature flexible detection method for waterproof roll based on image recognition in an embodiment of the application;

[0043] Figure 2 A structure diagram of a waterproof roll crack detection model in an embodiment of the application;

[0044] Figure 3 A structure diagram of a bending detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0045] Figure 4 A structure diagram of a backbone network of a bending detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0046] Figure 5 A structure diagram of a multi-scale feature fusion layer of a bending detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0047] Figure 6 A structure diagram of a bending detection head of a bending detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0048] Figure 7 A structure diagram of a crack detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0049] Figure 8 A flowchart of a frequency-space feature mixing module of a crack detection sub-module of a waterproof roll crack detection model in an embodiment of the application;

[0050] Figure 9A training flow diagram of the waterproofing membrane crack detection model in the embodiment of the present application is shown in the figure.

[0051] Figure 10 A structure diagram of the multi-level auxiliary information layer, auxiliary reversible branch and auxiliary detection head of the bending detection sub-module of the waterproofing membrane crack detection model in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present application will be specifically described below in conjunction with the accompanying drawings, wherein the drawings form a part of the present application and are used to explain the principles of the embodiments of the present application, but are not used to limit the scope of the present application.

[0053] One specific embodiment of the present application discloses a waterproofing membrane low-temperature flexibility detection method based on image recognition function, as shown in the figure, comprising the following steps S1-S3: Figure 1

[0054] Step S1, obtaining images of a plurality of waterproofing membrane samples subjected to low-temperature flexibility testing.

[0055] Specifically, after low-temperature flexibility testing of the waterproofing membrane, an industrial camera is used to shoot a plurality of waterproofing membrane samples to obtain images containing surface information of the waterproofing membrane.

[0056] Step S2, inputting each image into a trained waterproofing membrane crack detection model to detect cracks at the bending part, and obtaining crack information at the bending part of each sample; wherein the waterproofing membrane crack detection model comprises a bending detection sub-module and a crack detection sub-module; the bending detection sub-module is used to detect the bending part in the sample; and the crack detection sub-module is used to detect crack information of the bending part.

[0057] Specifically, the model structure of the waterproofing membrane crack detection model is as shown in the figure. Figure 2

[0058] For waterproofing membranes, cracks usually occur at the bending part of the waterproofing membrane. The bending part of the waterproofing membrane is first detected in the embodiment, and then the picture of the bending part is further judged for cracks through crack detection. This makes the crack detection more targeted, concentrates the detection resources on the bending part of the waterproofing membrane, can improve the detection accuracy of these areas, and at the same time reduces the detection in areas without cracks, thereby reducing the false detection caused by random texture or material unevenness.

[0059] Further, the structure of the bending detection sub-module is a YOLOv9 model structure, as shown in the figure, comprising an input layer, a backbone network, a multi-scale feature fusion layer and a bending detection head. Figure 3

[0060] ​​​Specifically, the input layer is configured to pre-process the image of the waterproofing membrane sample, and the pre-processing includes image denoising and image enhancement.

[0061] The image denoising uses a ZS-N2N denoising algorithm based on a convolutional neural network. Compared with traditional filtering denoising methods such as mean filtering algorithm and Gaussian filtering algorithm, the traditional methods have great limitations in processing complex noise, preserving image details and adapting to different scales of noise. The filtering denoising algorithm based on the convolutional neural network can automatically learn the noise characteristics, effectively model the noise, maximize the preservation of original image information, and improve the image quality.

[0062] To solve the problem of low contrast between the crack area and the background of the waterproofing membrane, the application adopts a histogram equalization image enhancement method to improve the image contrast. This method re-distributes the gray values of the original image, expands the gray values with small proportion in the image, and compresses the gray values with large proportion, so that the image gray values are uniformly distributed. After histogram equalization, the contrast between the crack target area and the background is higher, the details of the image are clearer, and the whole image has more semantic information, which will help to improve the accuracy of crack defect detection.

[0063] The backbone network, as shown in Figure 4 is configured to extract features from the pre-processed image and convert the image into a feature representation with rich semantic information. The Silence layer is configured to pass the input information to the backbone network. The Conv layer is a convolutional layer configured to extract local features of the image. The GELAN layer is a high-efficiency attention network layer configured to enhance the feature extraction capability and improve the accuracy of target detection. The Adown layer is a down-sampling layer configured to reduce the input vector by half and pass it to the next layer.

[0064] The multi-scale feature fusion layer, as shown in Figure 5 includes a small target detection layer, a medium target detection layer and a large target detection layer, each of which focuses on different size of detail information and is configured to fuse the features from the backbone network.

[0065] The bending detection head, as shown in Figure 6 includes a bounding box regressor (Box) and a classifier (Cls) configured to output a prediction box and a prediction label from the target detection layer of the multi-scale feature fusion layer. It should be noted that, in the embodiment, the multi-scale feature fusion layer of the bending detection sub-module includes three target detection layers, and therefore, the bending detection head of the bending detection sub-module also has three, which are configured to predict the features output by the target detection layers of different scales.

[0066] It should be noted that the detection frame output by the bending detection submodule is removed after the prediction frame with a confidence lower than 0.25 is removed, and then the detection frame with a higher overlap is removed through non-maximum suppression, and finally the label with the highest class confidence is selected as the classification label of the prediction frame to obtain the picture of the bending part of the waterproof roll.

[0067] Further, as shown in Figure 7 The crack detection submodule includes a high-frequency feature extraction module, a spatial feature extraction module, and a frequency-spatial feature mixing module.

[0068] The image of the bending part is input into the high-frequency feature extraction module and the spatial feature extraction module of the crack detection submodule, to obtain high-frequency features and RGB features of different scales of the image of the bending part.

[0069] After the high-frequency features and the RGB features of the same scale are fused in the frequency-spatial feature mixing module of the crack detection submodule, the crack information in the image of the bending part is output by the crack detection head.

[0070] Specifically, there is a lot of feature information in the frequency domain information of the waterproof roll crack image that is difficult to detect in the RGB space. The model relying only on the RGB spatial information of the waterproof roll image is far from enough. For example, when a crack appears on the waterproof roll, there will be a frequency transformation from a low-frequency signal of a flat image to a high-frequency signal of a sharp deformation in the vertical direction of the crack. It is difficult for the RGB space to perceive this signal transformation.

[0071] Further, the image of the bending part is sequentially input into the backbone network layer and the multi-scale feature fusion layer of the spatial feature extraction module to obtain RGB features of different scales; wherein the multi-scale feature fusion layer includes three target detection layers of different scales.

[0072] Specifically, the spatial feature extraction module uses a YOLOv9 model that removes the input layer and the detection head, and includes a backbone network and a multi-scale feature fusion layer.

[0073] Further, in the high-frequency feature extraction module, the image of the bending part is converted from spatial to frequency domain through discrete cosine transformation, and the high-frequency region weight of the picture is increased through a filter to obtain a high-frequency enhanced image. After the high-frequency enhanced image is restored to the spatial domain through inverse discrete cosine transformation, it is input into three convolution layers to obtain high-frequency features of three different scales.

[0074] Specifically, the discrete cosine transformation is a mathematical transformation for converting a signal from a spatial domain to a frequency domain, which can decompose an image into different frequency components. The filter is used to increase the weight of the high-frequency region of the picture and reduce the weight of the low-frequency region of the picture, so as to highlight the detailed features such as cracks in the image. The inverse discrete cosine transformation is the inverse process of the discrete cosine transformation, which is used to convert the processed frequency domain data back to the spatial domain.

[0075] It should be noted that the three high-frequency features of different scales obtained by the three convolutional layers correspond one-to-one to the three scale RGB features output by the spatial feature extraction module, and are used for feature fusion in the frequency-spatial feature mixing module. For example, the three scales are 160x160x128, 80x80x256 and 40x10x512, respectively.

[0076] Further, as shown in Figure 8 The high-frequency feature and the RGB feature are matrix multiplied to obtain a mixed frequency domain feature.

[0077] The high-frequency feature and the RGB feature are matrix multiplied to obtain a mixed frequency domain feature.

[0078] The high-frequency feature and the RGB feature are matrix multiplied to obtain a mixed frequency domain feature.

[0079] The high-frequency feature and the RGB feature are matrix added to obtain a high-frequency enhancement-RGB fusion feature.

[0080] The high-frequency feature and the RGB feature are matrix added to obtain a high-frequency enhancement-RGB fusion feature.

[0081] The high-frequency feature and the RGB feature are matrix added to obtain a high-frequency enhancement-RGB fusion feature.

[0082] Specifically, the frequency-spatial feature mixing module uses the high-frequency feature and the RGB feature to calculate the cross-attention weight, then uses the cross-attention matrix to enhance the attention features of the spatial and frequency domains, and finally combines the attention features of the spatial and frequency domains together and outputs.

[0083] It should be noted that the model structure of the crack detection head of the crack detection sub-module uses the detection head structure of the YOLOv9 model, removes the prediction boxes with a confidence lower than 0.25 in the output detection boxes, and then removes the detection boxes with a higher overlap degree through non-maximum suppression. Finally, the label with the highest class confidence is selected as the classification label of the prediction box, and the crack information of the waterproof coiled material is obtained.

[0084] Further, as shown in Figure 9 The waterproofing membrane crack detection model is obtained by training through the following method, including steps S201-S205:

[0085] Step S201, respectively constructing a bending detection training data set and a crack detection training data set; wherein the bending detection training data set includes sample images and labeled bending detection boxes; the crack detection training data set includes images in the labeled bending detection boxes and corresponding crack information.

[0086] Specifically, according to the size, position, shape and other characteristics of the waterproofing membrane crack, the LabelImg software is used to accurately mark the crack defects and the bending part of the membrane, draw the most accurate rectangular frame for the cracks and bending in all the pictures to be marked, and name the marked rectangular frame (true value frame) as crack and bending. Save all the marked rectangular frames in YOLO data format.

[0087] Step S202, loading the bending detection training data set to train the bending detection sub-module to obtain a trained bending detection sub-module.

[0088] Further, load the bending detection training data set, use the combination of the bounding box intersection over union loss function, the classification loss function and the confidence loss function, train the bending detection sub-module, use the stochastic gradient descent method to update the parameters of the bending detection sub-module by back propagation, and obtain the trained bending detection sub-module.

[0089] Specifically, as shown in Figure 10 The bending detection sub-module further includes a multi-level auxiliary information layer and an auxiliary reversible branch in the training; the multi-level auxiliary information layer is an integrated network inserted between the auxiliary reversible branch and the backbone network of the bending detection sub-module, used to receive gradient information of all target objects, aggregate the gradient information of all target objects in multiple levels of auxiliary information, and pass the gradient information to the backbone network of the bending detection sub-module for weight update; it should be noted that this module is only used in the training mode, because the returned gradient is obtained from the auxiliary reversible branch.

[0090] The structure of the auxiliary reversible branch is consistent with that of the backbone network, which is used to help the backbone network of the bending detection sub-module to better obtain rich return gradient information in the training process, and discard the branch in the inference process. Without losing accuracy, the detection speed of the model can be improved.

[0091] The output characteristics of the auxiliary reversible branch output the prediction box and the prediction label of the bending part through the auxiliary detection head, which are used to calculate the loss together with the prediction box and the prediction label of the bending part output by the bending detection head, so as to provide more gradient information in the training process and help to accelerate the convergence of the model.

[0092] Further, the boundary box intersection over union loss function uses a CIOU loss function, which is an improved version based on the original IOU (Intersection over Union) loss function, considers the complete intersection between target boxes, and introduces a correction factor to more accurately measure the similarity between target boxes.

[0093] The classification loss function uses a binary cross-entropy loss function.

[0094] The distribution focal loss function is a DFL (Distribution Focal Loss) loss function.

[0095] Specifically, in the training process, the boundary box intersection over union loss, the classification loss, and the distribution focal loss are added to obtain the total loss, and the total loss is optimized by backpropagation to optimize the parameters of the bending detection sub-module.

[0096] Step S203, pre-training the backbone network of the spatial feature extraction module of the crack detection sub-module based on contrastive learning to obtain a pre-trained backbone network.

[0097] Specifically, in order to avoid overfitting of the crack detection sub-module, the data set should include a large number of images under various conditions. However, for industrial scenes, the samples of waterproofing membranes are usually unbalanced, i.e. the number of normal waterproofing membrane samples is much larger than the number of waterproofing membrane crack samples. Therefore, the embodiment uses a contrastive learning-based method to pre-train the backbone network of the spatial feature extraction module of the crack detection sub-module, so that it learns the intra-class consistency and inter-class difference feature representation of the waterproofing membrane crack, lays a foundation for formal training of the model, improves the detection effect of the model, and reduces the occurrence of model overfitting.

[0098] Further, the pre-training includes the following steps:

[0099] A plurality of training data sets are constructed, each including a plurality of training sample pairs; wherein each training sample pair includes a positive sample and a negative sample; the positive sample in each training data set is a picture of the same crack region, and the negative sample is a different picture without a crack region.

[0100] Each negative sample in each training data set includes a plurality of crack-free region pictures of the same size as the positive sample and two pictures of the same size as the positive sample extracted on both sides of the long side of the positive sample.

[0101] Specifically, the crack of the waterproof roll material is usually a transverse rectangular crack, therefore, when constructing the negative sample pair of the sample, a picture with the same size as the positive sample is extracted from both sides of the long side of the positive sample, by selecting an area similar to the crack area but actually not containing the crack as the negative sample, the model can learn the subtle differences between the crack and the normal area, and enhance the difference of the features.

[0102] Further, for each training data set, load the training sample pair, use the contrast loss function to train the backbone network of the spatial feature extraction module of the crack detection sub-module, update the parameters of the backbone network of the spatial feature extraction module by gradient back propagation, and save the parameters of the backbone network of the spatial feature extraction module after the training is completed.

[0103] Specifically, during training, the positive sample is input into the decoder (the backbone network structure of the spatial feature extraction module) to obtain a positive sample feature vector matrix, the negative sample is input into the dynamic decoder (the backbone network structure for training in the spatial feature extraction module) to obtain a negative sample feature vector matrix, the contrast loss is calculated after the matrix multiplication of the positive sample feature vector matrix and the negative sample feature vector matrix, the gradient of the contrast loss is back propagated to the decoder, the parameters of the dynamic decoder are updated using the following formula after the parameters of the decoder are updated:

[0104] Dynamic decoder parameters = decoder parameters * 0.001 + previous dynamic decoder parameters * 0.999

[0105] It should be noted that the dynamic decoder parameters are updated synchronously using the above formula, which can be updated from the decoder parameters at a small scale, the model can slowly adapt to new knowledge, avoid performance fluctuations caused by large-scale updates, and allow the dynamic decoder to learn from the decoder, which may help to introduce feature diversity and improve the generalization ability of the model to different categories.

[0106] For example, the contrast loss function can use a temperature scaled cross-entropy loss function.

[0107] Step S204, load the crack detection training data set to train the crack detection sub-module containing the pre-trained backbone network to obtain a trained crack detection sub-module.

[0108] Further, load the crack detection training data set, use the combination of the bounding box small target detection loss function, the classification loss function and the distribution focal loss function to train the crack detection sub-module, update the parameters of the crack detection sub-module by random gradient descent method for back propagation to obtain a trained crack detection sub-module.

[0109] Specifically, in the waterproofing membrane crack image, there are many small crack images. When calculating the bounding box loss, the evaluation index based on the intersection over union is very sensitive to the position deviation of small objects, which greatly reduces the detection performance of the model. Therefore, the embodiment uses a bounding box small target detection loss function, i.e., a normalized Wasserstein distance (NWD) function, as an evaluation index to measure the similarity between two bounding boxes.

[0110] In training the crack detection sub-module, the classification loss function and the distribution focal loss function use the same loss function as the bending detection sub-module, i.e., the binary cross-entropy loss function and the DFL (Distribution Focal Loss) loss function, respectively.

[0111] In the training process, the bounding box small target detection loss, the classification loss, and the distribution focal loss are added to obtain the total loss, and the total loss is used to optimize the parameters of the crack detection sub-module through back propagation.

[0112] Step S205, based on the trained bending detection sub-module and the trained crack detection sub-module, a trained waterproofing membrane crack detection model is obtained.

[0113] Step S3, when the number of samples with cracks exceeds a preset value, it is determined that the low-temperature flexibility test is unqualified.

[0114] For example, 5 waterproofing membrane samples produced in the same batch and subjected to low-temperature flexibility tests are selected. When 2 or more samples are detected to have cracks, it is determined that the low-temperature flexibility test of the current batch of waterproofing membrane is unqualified.

[0115] In summary, the waterproofing membrane low-temperature flexibility detection method based on image recognition according to the embodiment of the application has the following beneficial effects:

[0116] 1. For waterproofing membranes, cracks usually occur at the bending part of the waterproofing membrane. The waterproofing membrane crack detection model of the application first uses the bending detection sub-model to detect the bending part of the waterproofing membrane, and then uses the crack detection sub-model to further determine whether there are cracks. Compared with directly detecting cracks on the waterproofing membrane image, it is more targeted and can improve the detection accuracy of these areas.

[0117] 2. The waterproofing membrane crack detection model uses the bending detection sub-model to detect the bending part of the waterproofing membrane, which can reduce the detection in areas without cracks, thereby reducing false positives caused by random textures or material non-uniformity, and reducing the image range that needs to be detected at high resolution, thereby saving computing resources and time.

[0118] 3. Since the samples of waterproof membranes are usually unbalanced in industrial scenarios, that is, the number of normal waterproof membrane samples is much greater than the number of samples of cracked waterproof membranes, this invention pre-trains the backbone network layer of the crack detection sub-model of the waterproof membrane crack detection model using a contrastive learning-based method. This allows the backbone network layer of the crack detection sub-model to learn the intra-class consistency and out-of-class difference features of waterproof membrane cracks, thereby improving the detection effect of the waterproof membrane crack detection model and reducing model overfitting.

[0119] 4. In constructing training sample pairs for the backbone network layer of the crack detection sub-model of the waterproof membrane crack detection model based on contrastive learning, the negative samples include two images of the same size as the positive samples extracted from both sides of the long side of the positive samples in the same training data group. By selecting areas similar to the crack area but not actually containing cracks as negative samples, the model can learn the subtle differences between crack and normal areas, enhance the difference of features, and more accurately identify crack areas.

[0120] 5. In the crack detection sub-model of the waterproof membrane crack detection model, the present invention extracts high-frequency feature information that is difficult to detect in the RGB space of the membrane image, which can better identify the existence of cracks, enhance the feature representation of small cracks, and thus improve the model's ability to detect small cracks.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An image recognition-based low-temperature flexibility detection method for waterproofing membranes, characterized in that, The method comprises the following steps: obtaining images of a plurality of waterproofing membrane samples subjected to low-temperature flexibility testing; inputting each image into a trained waterproofing membrane crack detection model to detect cracks at the bending part, and obtaining crack information of the bending part of each sample; wherein the waterproofing membrane crack detection model comprises a bending detection submodule and a crack detection submodule; the bending detection submodule is used to detect the bending part in the sample; the crack detection submodule is used to detect crack information of the bending part; the crack detection submodule comprises a high-frequency feature extraction module, a spatial feature extraction module, and a frequency-spatial feature mixing module; the image of the bending part is input into the high-frequency feature extraction module and the spatial feature extraction module of the crack detection submodule to obtain high-frequency features and RGB features of the image of the bending part at different scales; after the high-frequency features and the RGB features at the same scale are mixed in the frequency-spatial feature mixing module of the crack detection submodule, crack information in the image of the bending part is output by a crack detection head; when the number of samples with cracks exceeds a preset value, it is determined that the low-temperature flexibility test is unqualified.

2. The method of claim 1, wherein, the high-frequency features and the RGB features are mixed in the frequency-spatial feature mixing module of the crack detection submodule, comprising: the high-frequency features and the RGB features are multiplied in matrix to obtain mixed frequency domain features; the mixed frequency domain features are multiplied in matrix with the high-frequency features and the RGB features to obtain high-frequency enhanced features and RGB enhanced features; the high-frequency enhanced features and the RGB features are added in matrix to obtain high-frequency enhanced-RGB fusion features; the RGB enhanced features and the high-frequency features are added in matrix to obtain RGB enhanced-high-frequency fusion features; the high-frequency enhanced-RGB fusion features and the RGB enhanced-high-frequency fusion features are fused in matrix channel to obtain fusion features.

3. The method of claim 1, wherein, the image of the bending part in the image is sequentially input into a backbone network layer and a multi-scale feature fusion layer of the spatial feature extraction module to obtain RGB features at different scales; wherein the multi-scale feature fusion layer comprises three target detection layers at different scales.

4. The method of claim 1, wherein, in the high-frequency feature extraction module, the image of the bending part is converted from spatial domain to frequency domain by discrete cosine transformation, and the high-frequency region weight of the image is increased by a filter to obtain a high-frequency enhanced image; the high-frequency enhanced image is restored to spatial domain by inverse discrete cosine transformation, and is input into three convolution layers to obtain high-frequency features at three different scales.

5. The method of claim 1, wherein, the waterproofing membrane crack detection model is trained by the following method: a bending detection training data set and a crack detection training data set are constructed respectively; wherein the bending detection training data set comprises sample images and labeled bending detection boxes; the crack detection training data set comprises images in the labeled bending detection boxes and corresponding crack information; the bending detection submodule is trained by loading the bending detection training data set to obtain a trained bending detection submodule; Pre-training the backbone network of the spatial feature extraction module of the crack detection sub-module based on contrastive learning to obtain a pre-trained backbone network; Loading the crack detection training data set to train the crack detection sub-module containing the pre-trained backbone network to obtain a trained crack detection sub-module; Based on the trained crack detection sub-module and the trained crack detection sub-module, a trained waterproof roll crack detection model is obtained.

6. The method of claim 5, wherein, The pre-training of the backbone network of the spatial feature extraction module of the crack detection sub-module based on contrastive learning to obtain a crack detection basic backbone network comprises: A plurality of training data sets are constructed, each training data set comprising a plurality of training sample pairs; wherein the training sample pair comprises a positive sample and a negative sample; the positive sample in each training data set is a picture of the same crack region, and the negative sample is a different picture without a crack region; For each training data set, load the training sample pair, use the contrastive loss function to train the backbone network of the spatial feature extraction module of the crack detection sub-module, update the backbone network parameters of the spatial feature extraction module using gradient backpropagation, and save the backbone network parameters of the spatial feature extraction module after training.

7. The method of claim 6, wherein, The negative sample in each training data set includes a plurality of crack-free region pictures of the same size as the positive sample and two pictures of the same size as the positive sample extracted on both sides of the long side of the positive sample.

8. The method of claim 5, wherein, The loading of the crack detection training data set to train the crack detection sub-module to obtain a trained crack detection sub-module comprises: loading the crack detection training data set, using a combination of the bounding box intersection over union loss function, the classification loss function and the distribution focal loss function, training the crack detection sub-module, updating the parameters of the crack detection sub-module using the stochastic gradient descent method for backpropagation, and obtaining the trained crack detection sub-module. The loading of the crack detection training data set to train the crack detection sub-module containing the pre-trained backbone network to obtain a trained crack detection sub-module comprises: loading the crack detection training data set, using a combination of the bounding box small target detection loss function, the classification loss function and the distribution focal loss function, training the crack detection sub-module, updating the parameters of the crack detection sub-module using the stochastic gradient descent method for backpropagation, and obtaining the trained crack detection sub-module.

9. The method according to any one of claims 1 to 8, characterized in that, The structure of the crack detection sub-module is a YOLOv9 model structure.

Images