High-quality cable bent tower mass concrete image generation method and quality detection method

By using an improved StyleGAN-XL generative adversarial network and Max-DeepLab segmentation network, the problem of scarce image data for large-volume concrete structures of Sota was solved, enabling high-quality image generation and real-time accurate analysis of crack detection, thus improving the performance and accuracy of the detection model.

CN119048938APending Publication Date: 2024-11-29EAST CHINA JIAOTONG UNIVERSITY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411118037.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Due to their large size and complexity, it is difficult to obtain high-quality, all-around image data for large-volume concrete structures like the Sota. Traditional methods cannot effectively monitor and analyze cracks, resulting in low detection accuracy.

Method used

An improved StyleGAN-XL generative adversarial network is used to expand and enhance images captured by UAVs, generating high-quality images of large-volume concrete structures on cable towers. An improved Max-DeepLab segmentation network is used for crack detection. Combined with UAV image acquisition, data annotation, and image segmentation modules, real-time quality monitoring is achieved.

Benefits of technology

It generates a large number of high-quality images that are similar to the actual environment, enriches the dataset, improves the accuracy and reliability of crack detection, can accurately identify and locate cracks, and achieves real-time, high-accuracy quantitative analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048938A_ABST
    Figure CN119048938A_ABST
Patent Text Reader

Abstract

The invention relates to a high-quality cable bent tower mass concrete image generation method and a quality detection method, and the generation method comprises the following steps: carrying out the high-altitude shooting of a cable bent tower through an unmanned aerial vehicle, obtaining an image of cable bent tower concrete, and marking the position and range of a crack in the image; and expanding and enhancing the obtained image by using an improved StyleGAN-XL generative adversarial network to obtain a high-quality cable bent tower mass concrete image. Virtual images are obtained through the generation method, and a large number of virtual images serve as a target data set R of a segmentation network; and training an improved MAX-DeepLab segmentation network by using the target data set R, wherein the improved MAX-DeepLab segmentation network is used for detecting the crack width of mass concrete. According to the method, a high-quality cable bent tower mass concrete image can be obtained, capacity expansion and enhancement processing can be carried out on a small number of unmanned aerial vehicle shot images, and the capacity of a segmentation model in the aspect of fine crack detection is improved by utilizing a generated high-quality data capacity expansion and enhancement data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of image semantic segmentation and intelligent construction technology, and in particular to a high-quality cable tower mass concrete image generation method, quality detection method and system. Background Art

[0002] The cable tower is a key load-bearing structure in a suspension bridge or cable-stayed bridge, responsible for supporting the main cables or stay cables of the bridge, and thus bearing the load of the bridge deck. The cable tower is usually located at both ends of the bridge or on the main pier in the middle of the bridge. Its design and construction are crucial to the stability and safety of the entire bridge.

[0003] Cable tower mass concrete usually refers to the mass concrete used for the foundation, tower body and other structures of bridge cable towers. This type of concrete construction has high requirements for materials, construction technology and testing technology due to its large volume and complex structure.

[0004] The construction of large-volume concrete towers faces many technical challenges, especially when monitoring and analyzing the state of concrete, the acquisition and processing of image data is particularly complex. This is due to the large size and complex shape of the tower structure, and the need to deal with the unique thermodynamic problems of large-volume concrete.

[0005] During the hardening process of large-volume concrete, a large amount of heat is generated due to the cement hydration reaction, which causes the internal temperature of the concrete to rise significantly. Due to the large volume of concrete, the internal heat is difficult to dissipate quickly, resulting in a significant temperature gradient between the inside and the surface. This temperature gradient will generate thermal stress during the cooling process. If the thermal stress exceeds the tensile strength of the concrete, it is easy to cause cracks. The existence of the temperature gradient causes the concrete structure to be subjected to uneven stress distribution during the cooling process, especially between the inside and the surface of the concrete. This uneven stress distribution easily causes micro cracks in the concrete, which expand into macro cracks over time. Since cracks affect the durability and bearing capacity of concrete, it is crucial to effectively monitor cracks in large-volume concrete. Since cracks in large-volume concrete may cover a wide area and the cracks are small and scattered, high-pixel images are required for deep learning to capture the details of the cracks. Conventional concrete crack images may not meet the requirements of crack recognition network training for large-volume concrete.

[0006] In addition, the towers are usually very tall and have complex surfaces, so shooting of different parts needs to be done at high altitudes. Traditional ground shooting methods cannot cover all surfaces of the tower, which makes it difficult to rely on ground shooting data for crack analysis. If drones are used to obtain image data of the entire tower, operating drones for precise photography requires professional technology and equipment, and factors such as weather and flight safety must also be considered, which makes data acquisition difficult and results in very little data being available.

[0007] Therefore, due to the size and complexity of the cable tower structure, it is a major problem to fully obtain high-quality image data. Existing generative networks cannot simply complete the expansion and enhancement of high-quality images. The present invention proposes a high-quality cable tower mass concrete image generation method, quality detection method and system. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the above technologies and provide an intelligent, scalable, high-quality cable tower mass concrete image generation method, quality detection method and system.

[0009] In order to solve the above technical problems, the technical solution of the present invention is:

[0010] In a first aspect, the present invention provides a method for generating a high-quality mass concrete image of a cable tower, the method comprising the following contents:

[0011] Use drones to take aerial photos of the cable tower, obtain images of the cable tower concrete, and mark the locations and ranges of cracks in the images;

[0012] The improved StyleGAN-XL generative adversarial network is used to expand and enhance the obtained images to obtain high-quality large-volume concrete images of cable towers;

[0013] The improved StyleGAN-XL generative adversarial network adopts a deep learning GAN network framework, including generator, discriminator and classifier modules:

[0014] The generator is used to generate images of different resolutions, including an embedding module, a latent code, a mapping network module, a style code, a Fourier feature change, a dilated convolution module, a multi-layer retuned convolution module of different resolutions, and a TORGB module; the input image is preprocessed by the embedding module, and then enters the mapping network module together with the latent code, and the class label obtained by the classifier module is also input into the mapping network module to obtain the style code W, the style code enters the dilated convolution module after the Fourier feature change for feature extraction, and then enters the retuned convolution module of each layer for splicing, and then generates images of various resolutions through the TORGB module; if the resolution is less than 20482 When , it means that the resolution requirement is not met, and then it enters the discriminator;

[0015] The discriminator includes two paths, the upsampling path is composed of upsampling, multi-head dilated attention convolution module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the downsampling path is composed of downsampling, mixed attention visual perception module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the four independent discriminators of the two paths are respectively subjected to adversarial learning with the image features generated by the generator that do not meet the resolution requirements after being processed by the embedding module to generate high-quality images;

[0016] Among the images that do not meet the resolution requirements, the relatively high-resolution images enter the upsampling path, and the relatively low-resolution images enter the downsampling path. After continuous confrontation between the two different paths of upsampling and downsampling, high-quality images are generated. The generated images that do not meet the resolution requirements but are of high quality are finally sent to the classifier module for the generator to continue to generate images that better meet the requirements;

[0017] The hybrid attention visual perception module adopts a VIT module, and the multi-head dilated attention convolution module includes multiple attention heads, and a dilated convolution is set in each attention head, and the dilation rate of each dilated convolution is different;

[0018] The retuning convolution module includes a modulation module, a demodulation module, a 3X3 convolution and a leaky Relu function, a double upsampling, a double tangent activation function, a double downsampling and a cropping image module. The generated style learned from the style code is affine transformed with the upper layer output feature, and then after being operated by the modulation module and the demodulation module, it enters the 3X3 convolution together with the upper layer output feature processed by the exponential moving average operation and is processed by the leaky Relu function. The image obtained after the double upsampling, the double tangent activation function and the double downsampling is cropped by the cropping image module to obtain the output feature map Q of the retuning convolution module;

[0019] The TORGB module includes a modulation module and a 3X3 air-conditioning convolution, the number of holes is 2, and the feature map Q undergoes affine transformation and modulation module operation in the TORGB module, and enters the 3X3 air-conditioning convolution for processing together with the upper-layer output features processed by the exponential moving average operation.

[0020] Furthermore, in the generator, the input image is first preprocessed by the embedding module, and the shallow features obtained are concatenated with a suitable latent code Z selected or interpolated from the pre-trained latent space and then enter the mapping network module for splicing. At the same time, the class label C obtained by the classifier module is concatenated by the mapping network module and then mapped into a style code W; the style code W is concatenated with the Fourier feature change and then enters the dilated convolution module for feature re-extraction;

[0021] After that, the 8×8 reconvolution module outputs an 8×8 image. At the same time, the 8×8 image is spliced ​​with W and then outputted through the 16×16 reconvolution module to generate a 16×16 image. At the same time, the 16×16 image is spliced ​​with W and then outputted into the 32×32 reconvolution module. After continuous iteration, the 1024×1024 image is spliced ​​with W and then outputted through the 2048×2048 reconvolution module to generate a 2048×2048 high-resolution image. Images of different resolutions are outputted through the TORGB module. If the resolution of the image is less than 2048, 2 Then enter the discriminator to operate.

[0022] Furthermore, the acquired images are used to train the improved StyleGAN-XL generative adversarial network. During the training process, the number of network layers of the generator and the discriminator is gradually increased, so that the network first learns large-scale structural information and then gradually shifts its attention to small-scale detail learning.

[0023] At the beginning, 11 layers are used for progressive growth at a resolution of 16X16. Initially, 11 16×16 retuned convolution modules are set in series, and each retuned convolution module is a layer. Each time the resolution increases, 2 layers are cut off and 7 new layers are added. For the final stage of 2048X2048, only 5 layers are added without subtraction. There are 44 layers from the first 16×16 retuned convolution module to the last 2048×2048 retuned convolution module, and the maximum resolution is 2048X2048. A fixed growth plan is used to train each stage until the similarity index FID stops decreasing. After adding new layers, the existing layers with lower resolution will remain fixed and only the latest layers will be trained. The model is trained on V100 for 24 hours at a resolution of 2048X2048 and evaluated using the Fid evaluation indicator. Training is terminated when Fid is below 5. The trained improved StyleGAN-XL generative network is used to generate a large number of high-quality virtual images with a resolution of 2048X2048.

[0024] In a second aspect, the present invention provides a method for detecting the image quality of a cable tower mass concrete, wherein the method uses the generation method described in any one of claims 1 to 3 to obtain a virtual image, and uses a large number of virtual images as a target data set R of a segmentation network;

[0025] The segmentation network adopts an improved MAX-DeepLab segmentation network, and the improved MAX-DeepLab segmentation network is trained using a target data set R for large-volume concrete crack width detection.

[0026] Furthermore, the entire network of the improved Max-DeepLab segmentation network adopts a Transformer framework, including an encoder and a decoder;

[0027] The encoder consists of a preprocessing module, a 3-layer linear bottleneck module, a 4-layer linear bottleneck module, a 6-layer axial attention bottleneck module, and a dual-path Transformer module;

[0028] The input image is processed by the preprocessing module, and then passes through a 3-layer linear bottleneck module and a 4-layer linear bottleneck module to extract features to obtain a pixel feature map A. The pixel feature map A enters a 6-layer axial attention bottleneck module and then enters the pixel path of the dual-path Transformer module. The dual-path Transformer module also inputs a global memory, which is a trainable parameter.

[0029] The decoder includes a plurality of stacked decoder bottleneck modules, each of which includes four 1X1 convolutions and a depth-separable convolution module. On the pixel path, each stacked decoder bottleneck module has two inputs, one is the feature output obtained in the previous layer, and the other is the pixel feature map A output by the 4-layer linear bottleneck module.

[0030] The output of multiple stacked decoder bottleneck modules is the decoding feature E. After being processed by a 5X5 convolution module and a 1X1 convolution module, the decoding feature E is matrix multiplied with the mask features extracted by two fully connected layers on the global path to obtain a mask image; finally, after passing through a batch normalization module, the final mask image is output for quantitative measurement of cracks.

[0031] Furthermore, the preprocessing module in the encoder includes a 7×7 standardized convolution, a group normalization, a ReLU activation function and a pooling layer connected in sequence;

[0032] The linear bottleneck module has three built-in 1X1 convolutions, one 3X3 convolution, and four batch normalization processing modules; the input feature map is sequentially processed by 1X1 convolution dimension increase and batch normalization, 3X3 depth convolution and batch normalization, 1X1 convolution dimension reduction and batch normalization to obtain a fused feature map P. The first two convolution activation functions use the Relu function, and the last dimension reduction convolution uses a linear activation function instead of the Relu function; the input image is again subjected to 1X1 convolution and batch normalization, and matrix addition is performed with the obtained feature map P, and after a linear activation function, an output feature map of a layer of linear bottleneck module is obtained.

[0033] Furthermore, in the training process of the improved MAX-DeepLab segmentation network, 32 TPU cores were used to train the model, the number of iterations was 100k, 120 epochs, the batch size was 64, the Radam and Lookahead optimization algorithms were used, and the scheduling learning rate was 10 -3 , the backbone learning rate multiplier is 0.1, and the weight decay rate is 10 -4 , the descent path rate is 0.2, and the training model is terminated when it reaches convergence.

[0034] In a third aspect, the present invention provides a cable tower mass concrete image quality detection system, the detection system comprising:

[0035] The image collection and processing module uses a drone to take aerial photos of the cable tower, obtains multiple frames of images including different parts of the cable tower, and pre-processes the images;

[0036] A data annotation module is used to annotate the preprocessed images, annotate the large-volume concrete cracks in the preprocessed images, and distinguish their locations and ranges;

[0037] The data expansion and enhancement module is connected to the data annotation module, and uses the annotated images to train the improved StyleGAN-XL generative adversarial network to obtain a trained improved StyleGAN-XL generative adversarial network to generate sufficient number and high-quality virtual images;

[0038] The image segmentation module is connected with the data annotation module and the data expansion and enhancement module to obtain the annotated image and the generated virtual image, so as to train the improved Max-DeepLab segmentation network for detecting and quantifying the crack width of the massive concrete of the cable tower;

[0039] The early warning feedback module is connected to the image segmentation module. The image segmentation module detects the width of the cracks in the massive concrete of the cable tower, and promptly warns and feeds back the situation of the cracks in the massive concrete of the cable tower to the background system. The background system automatically issues an alarm and prompts relevant personnel to take measures.

[0040] Furthermore, when checking the width of cracks in mass concrete, if the width of the cracks is less than 0.30 mm, the quality of the mass concrete is considered qualified, and if the width of the cracks is greater than 0.30 mm, it is considered that an alarm check is required.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention aims at the fact that the massive concrete structures of cable towers are usually located at high altitudes, have complex shapes, and it is difficult to obtain all-round high-quality images during construction or maintenance, and data is scarce. Therefore, a method for obtaining high-quality massive concrete images of cable towers is proposed, which can expand and enhance a small number of images taken by drones. The data sets expanded and enhanced using the generated high-quality data can supplement the rare crack patterns that were originally difficult to obtain in the data set, provide more samples for model learning, and improve the model's ability in detecting fine cracks, so as to train a massive concrete crack detection model for cable towers with better performance and stronger generalization ability. The rich data samples enable the model to better learn the characteristics of different cracks, thereby improving the accuracy and reliability of detection.

[0043] The present invention is based on high-quality generation. When faced with complex cable tower mass concrete structures, the segmentation model can more effectively identify and locate cracks and accurately analyze the shape and size of the cracks.

[0044] The method of the present invention can generate a large number of large-volume concrete images similar to the actual environment. These generated images can not only increase the scale of the data set, but also cover more crack types and morphologies, thereby enhancing the diversity of the data set.

[0045] The improved Max-DeepLab is used as the segmentation network model in the detection method of the present invention, which greatly improves the accuracy of recognition and quantification, enables the system to achieve real-time high-accuracy quantification, and extracts features from both shallow features and deep features. While ensuring the quality of feature information, it also solves the problem of the influence of shallow features on deep features caused by the network layer being too deep, making the model more stable and reliable. The improved Max-DeepLab can effectively reduce the amount of calculation and the number of parameters, while improving the efficiency and speed of the model. This structure can significantly reduce the size and computing requirements of the model while maintaining relatively high performance, especially in resource-constrained environments such as mobile devices. It is easier to implement and deploy on different platforms, and can streamline the calculation model, thereby achieving high-precision calculations on smaller devices, and improving the accuracy of mass concrete quality detection of cable towers.

[0046] In summary, the present invention adopts StyleGAN-MAX to continuously generate richer and higher-quality data sets to realize automatic learning function, and combines the improved Max-DeepLab segmentation network to realize high-precision rapid monitoring of large-volume concrete cracks. It can accurately and efficiently detect micro crack defects and quantify them, providing an effective means to ensure the quality of cable towers, and can extract the specific morphology and distribution of large-volume concrete cracks, realize real-time quality monitoring and quantitative analysis of cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the improved StyleGAN-XL generative adversarial network structure in the present invention;

[0048] Figure 2 It is a schematic diagram of the structure of the improved StyleGAN-XL generative adversarial network convolution module in the present invention;

[0049] Figure 3 It is a schematic diagram of the structure of the improved StyleGAN-XL generative adversarial network TORGB module in the present invention;

[0050] Figure 4 It is a schematic diagram of the structure of the improved Max-DeepLab segmentation network in the present invention;

[0051] Figure 5 It is a schematic diagram of the structure of the preprocessing module in the improved Max-DeepLab segmentation network of the present invention;

[0052] Figure 6 Schematic diagram of the structure of the linear bottleneck module in the improved Max-DeepLab segmentation network of the present invention;

[0053] Figure 7 Schematic diagram of the structure of the dual-path converter module in the improved Max-DeepLab segmentation network of the present invention;

[0054] Figure 8 It is a flow chart of a method for detecting image quality of mass concrete of a cable tower according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention is further explained below in conjunction with the embodiments and drawings, but this is not intended to limit the scope of protection of the present application.

[0056] Example 1

[0057] The high-quality cable tower mass concrete image generation method of this embodiment includes the following contents:

[0058] Use an unmanned aerial vehicle (UAV) to take aerial photos of the cable tower, obtain images of the cable tower concrete, and mark the location and range of cracks in the images;

[0059] The improved StyleGAN-XL generative adversarial network is used to expand and enhance the obtained images to obtain high-quality large-volume concrete images of cable towers;

[0060] The improved StyleGAN-XL generative adversarial network adopts a deep learning GAN network framework, represented by StyleGAN-MAX, and its overall structure is as follows: Figure 1 As shown, it includes generator, discriminator and classifier modules:

[0061] The generator is used to generate images of different resolutions, including an embedding module, a latent code, a mapping network module, a style code, a Fourier feature change, a dilated convolution module, a multi-layer retuned convolution module of different resolutions, and a TORGB module; the input image is preprocessed by the embedding module, and then enters the mapping network module together with the latent code, and the class label obtained by the classifier module is also input into the mapping network module to obtain the style code W, the style code enters the dilated convolution module after the Fourier feature change for feature extraction, and then enters the retuned convolution module of each layer for splicing, and then generates images of various resolutions through the TORGB module; if the resolution is less than 2048 2 When , it means that the resolution requirement is not met, and then it enters the discriminator;

[0062] The discriminator includes two paths, the upsampling path is composed of upsampling, multi-head dilated attention convolution module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the downsampling path is composed of downsampling, mixed attention visual perception module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the four independent discriminators of the two paths are respectively subjected to adversarial learning with the image features generated by the generator that do not meet the resolution requirements after being processed by the embedding module to generate high-quality images;

[0063] Among the images that do not meet the resolution requirements, the relatively high-resolution images enter the upsampling path, and the relatively low-resolution images enter the downsampling path. After continuous confrontation between the two different paths of upsampling and downsampling, high-quality images are generated. The generated images that do not meet the resolution requirements but are of high quality are finally sent to the classifier module for the generator to continue to generate images that better meet the requirements;

[0064] The hybrid attention visual perception module adopts a VIT module, and the multi-head dilated attention convolution module includes multiple attention heads, and a dilated convolution is set in each attention head, and the dilation rate of each dilated convolution is different;

[0065] The retuning convolution module includes a modulation module, a demodulation module, a 3X3 convolution and a leaky Relu function, a double upsampling, a double tangent activation function, a double downsampling and a cropping image module. The generated style learned from the style code is affine transformed with the upper layer output feature, and then after being operated by the modulation module and the demodulation module, it enters the 3X3 convolution together with the upper layer output feature processed by the exponential moving average operation and is processed by the leaky Relu function. The image obtained after the double upsampling, the double tangent activation function and the double downsampling is cropped by the cropping image module to obtain the output feature map Q of the retuning convolution module;

[0066] The TORGB module includes a modulation module and a 3X3 air-conditioning convolution, the number of holes is 2, and the feature map Q undergoes affine transformation and modulation module operation in the TORGB module, and enters the 3X3 air-conditioning convolution for processing together with the upper-layer output features processed by the exponential moving average operation.

[0067] When annotating the preprocessed image in the present invention, the crack is framed with a rectangular frame in the image to mark its position and range. The improved StyleGAN-XL is used for expansion and enhancement processing, and more abundant and high-quality virtual data are continuously generated to realize the automatic learning function, obtain the desired image data, and construct the target data set.

[0068] Example 2

[0069] The improved StyleGAN-XL generative adversarial network modules in this embodiment are introduced as follows:

[0070] In the generator, the input image is first preprocessed by the embedding module, and the shallow features obtained are joined with a suitable latent code Z selected or interpolated from the pre-trained latent space and then enter the mapping network module for splicing. At the same time, the class label C obtained by the classifier module is spliced ​​into a style code W after being spliced ​​by the mapping network module. The style code W is spliced ​​with the Fourier feature change and then enters the dilated convolution module for feature re-extraction. By inserting blanks between the elements of the standard convolution kernel, the receptive field can be expanded without adding additional computational burden. While expanding the receptive field, the dilated convolution does not add additional parameters, so that the model maintains high efficiency while increasing the expressive power. The network is allowed to capture information from fine-grained to coarse-grained at different levels, which enhances the ability to capture features of different scales, and enables the images generated later to have more sufficient feature information and prevent the omission of feature information.

[0071] After that, the 8×8 reconvolution module will generate an 8×8 image. At the same time, the 8×8 image will be spliced ​​with W and then output through the 16×16 reconvolution module to generate a 16×16 image. At the same time, the 16×16 image will be spliced ​​with W and then output through the 32×32 reconvolution module. After continuous iteration, the 1024×1024 image will be spliced ​​with W and then output through the 2048×2048 reconvolution module to generate a 2048×2048 high-resolution image. Images of different resolutions are output through the TORGB module. If the resolution of the image is less than 2048 2 Then enter the discriminator for operation.

[0072] The convolution module is adjusted as follows Figure 2 As shown in the figure, a modulation module, a demodulation module, a 3X3 convolution and a Leaky Relu function, a double upsampling, a double tangent activation function, a double downsampling and a cropped image (CROP) module are built in. The generated style learned from the style code is affine transformed with the upper layer output features, and then after the modulation module and the demodulation module, it enters the 3X3 convolution together with the upper layer output features processed by the exponential moving average operation and is processed by the Leaky Relu function. After the double upsampling, double tangent activation function and double downsampling, the image obtained is cropped by the CROP module to obtain a suitable feature map Q. The main function of the modulation and demodulation modules is to break the original AdaIN into explicit normalization. Both modules are performed on the mean and standard deviation of each feature map, based on the expected statistics of the feature map, and the original normalization is removed while retaining controllability. The Leaky Relu activation function can give the input a very small slope when the input is negative, ensure the activation state of the neuron, and expand the perception range of the input. For image quality processing, the consistency between input and output is ensured in terms of image color, contrast, and brightness. In addition, the bi-tangent activation function is used to convert the obtained features into nonlinearity, and the image quality is improved after passing through the TORGB module.

[0073] TORGB module such as Figure 3 As shown in the figure, a modulation module and a 3X3 air-conditioning convolution are built in, and the number of holes is 2. The feature map Q is processed by the affine transformation and modulation module operation in the TORGB module together with the upper layer output features processed by the exponential moving average operation, and then enters the 3X3 air-conditioning convolution for processing. This module uses 3X3 hole convolution here, which not only increases the size of the receptive field, but also reduces the information loss of the image. The effect is more obvious than that of traditional convolution, and high-resolution image quality can be obtained.

[0074] In the discriminator, the resolution is less than 2048 2The generated image will enter the discriminator. The discriminator includes two paths, one is the upsampling path and the other is the downsampling path. The upsampling path receives images with relatively high resolution, and the downsampling path receives images with relatively low resolution. The two paths are the same except that the multi-head dilated attention convolution module is replaced by the mixed attention visual perception module. The multi-head dilated attention convolution module receives the upsampled image, performs feature fusion through the 1X1 convolution module and the random 3X3 convolution layer, and then inputs the obtained feature map into four independent discriminators for judgment after bilinear upsampling. On the other path, the mixed attention visual perception module receives the downsampled image to keep the memory requirement low, but still use its global feedback. The feature fusion is performed through the 1X1 convolution module and the random 3X3 convolution layer, and the obtained feature map is also input into four independent discriminators for judgment after bilinear upsampling. The image features generated by the generator will also be processed by the embedding module. The features processed by the embedding module will enter the eight independent discriminators for confrontation by the independent discriminators, so that the generator can continuously generate more and higher quality images. The image output from the generator that meets the resolution requirements will also be sent to the classifier module CLF for classifier guidance, and the generated class label C will enter the mapping network module for training.

[0075] The introduced multi-head dilated attention convolution module is a multi-head attention plus dilated convolution, setting a different dilation rate for each attention head. When calculating the attention matrix, the multi-head attention mechanism splits the input tensor X into h sub-tensors, each of which is the attention information learned in a different way. Then, for each sub-tensor, a self-attention calculation is performed to obtain an output tensor O i Finally, the h output tensors are concatenated together to obtain the final output tensor O.

[0076] Specifically, let X∈R nxd is the input tensor, Q∈R dxd , K∈R dxd , V∈R dxd are the learned d-dimensional query, key, and value vectors respectively, and h is the number of heads. Specifically, assuming the input sequence is x1, x2, ..., x n , then the calculation of multi-head attention is as follows:

[0077] MultiHead(X)=Concat(head1,...,head h )W O

[0078] Among them, Q, K, and V represent the queries, keys, and values ​​of the input sequence respectively. i represents the i-th attention head,

[0079] W O is the weight parameter of linear projection. Each attention head is obtained by performing self-attention calculation on Q, K, and V. The calculation formula is as follows:

[0080]

[0081] Among them, d K is the dimension size of Q, K, which is used to scale the size of attention.

[0082] Head i =Attention(XQ i ,XK i ,XV i ), i=1,...,h

[0083]

[0084] In the above formula, W Q ∈R dxhq , W K ∈R dxhk , W V ∈R dxhv and W O ∈R hdxd are the corresponding learned projection matrices, q, k, v are the dimensions of the query, key, and value vectors, respectively.

[0085] The advantage of the multi-head attention mechanism is that it can handle multiple focus issues and can better handle complex semantic relationships. The dilated convolution is introduced into the multi-head attention module, and a different dilation rate is set for each attention head to make the output denser. The introduction of the multi-head dilated attention convolution module here makes the output information of relatively high-resolution images richer after passing through the module, adapting to the discriminator task without reducing the resolution, and can continuously generate better quality images for the generator.

[0086] The improved StyleGAN-XL generative adversarial network is trained using the acquired images. During the training process, the number of network layers of the generator and the discriminator is gradually increased, which allows the network to first learn large-scale structural information and then gradually shift its attention to small-scale detail learning; instead of learning information at all scales at the same time, the output resolution of each stage is doubled. Initially, 11 layers are used for progressive growth at a resolution of 16X16 (that is, 11 16×16 retuned convolution modules are initially set in series, and each retuned convolution module is a layer). Each time the resolution increases, 2 layers are cut off and 7 new layers are added. For the last stage of 2048X2048, only 5 layers are added because the last two layers are not discarded. This is equivalent to a total of 44 layers from the first 16×16 retuned convolution module to the last 2048×2048 retuned convolution module, with a maximum resolution of 2048X2048. There is no fixed growth plan in between, but each stage needs to be trained until the similarity index FID stops decreasing. After adding new layers, the layers with lower resolutions will remain fixed to prevent mode collapse, and only the latest layers will be trained while keeping other layers unchanged. The training is carried out on V100 at a resolution of 2048X2048 for 24 hours, and the Fid evaluation indicator is used to evaluate it. When Fid is below 5, the training is terminated; the improved StyleGAN-XL generative network after training is used to generate 10,000 virtual images with a resolution of 2048X2048 to form virtual data;

[0087] The target dataset R for training the improved Max-DeepLab segmentation network is constructed using virtual data. The target dataset R is divided into a training set plus a validation set and a test set in a ratio of 7:3 for training and validation of the segmentation network.

[0088] Example 3

[0089] This embodiment is configured on the basis of Embodiment 2. The entire network of the improved Max-DeepLab segmentation network adopts the Transformer framework, and its overall structure is as follows: Figure 4 As shown, it includes an encoder and a decoder;

[0090] The encoder consists of a preprocessing module, a 3-layer linear bottleneck module, a 4-layer linear bottleneck module, a 6-layer axial attention bottleneck module, and a dual-path Transformer module;

[0091] The input image is processed by the preprocessing module, and then passes through a 3-layer linear bottleneck module and a 4-layer linear bottleneck module to extract features to obtain a pixel feature map A. The pixel feature map A enters a 6-layer axial attention bottleneck module and then enters the pixel path of the dual-path Transformer module. The dual-path Transformer module also inputs a global memory, which is a trainable parameter.

[0092] The decoder includes a plurality of stacked decoder bottleneck modules, each of which includes four 1X1 convolutions and a depth-separable convolution module. On the pixel path, each stacked decoder bottleneck module has two inputs, one is the feature output obtained in the previous layer, and the other is the pixel feature map A output by the 4-layer linear bottleneck module.

[0093] On the one hand, the feature output obtained in the previous layer is subjected to a 1X1 convolution for feature fusion and then up-sampled to obtain a new feature B; on the other hand, the feature output obtained in the previous layer is subjected to a 1X1 convolution to obtain a new feature D;

[0094] The pixel feature map A undergoes a 1X1 convolution for feature fusion, and then upsamples to obtain new features, and then undergoes a depth-wise separable convolution for feature fusion to obtain a new feature C.

[0095] The new features B, C, and D are added together in a matrix and then fused through a 1X1 convolution. This process is repeated again, and features of different scales are extracted in multiple stacked decoder bottleneck modules. The decoded feature E is then fused through bilinear adjustment and simple summation.

[0096] The decoded feature E is processed by a 5X5 convolution module and a 1X1 convolution module, and then matrix multiplied with the mask feature extracted by two fully connected layers on the global path to obtain a mask image; finally, the final mask image is output after passing through a batch normalization module to achieve quantitative measurement of cracks;

[0097] Next, each module of the improved Max-DeepLab is further introduced;

[0098] In the encoder, the collected images are first input into the preprocessing module for processing. The input features are respectively extracted through a 3-layer linear bottleneck module and a 4-layer linear bottleneck module. The pixel features obtained after extraction enter a 6-layer axial attention bottleneck module. The pixel features obtained after calculation and a global memory (actually a trainable parameter) of the input enter the dual-path Transformer module.

[0099] The present invention introduces a preprocessing module into the network, such as Figure 5 As shown, the preprocessing module includes 7×7 standardized convolution, group normalization, Relu activation function and pooling layer connected in sequence, which can solve the gradient explosion problem and prevent data overfitting, has stable training dynamics, does not depend on the statistics between batches, does not depend on the batch size, is insensitive to the batch size, is particularly suitable for situations where the batch size is limited, improves the performance on small data sets, simplifies the deployment and maintenance of neural networks, and can accelerate the training efficiency of neural networks.

[0100] The linear bottleneck module is as follows Figure 6 As shown, three 1X1 convolutions, one 3X3 convolution, and four batch normalization processing modules are built in. The input feature map is processed by 1X1 convolution dimension increase and batch normalization, 3X3 depth convolution and batch normalization, 1X1 convolution dimension reduction and batch normalization to obtain a fused feature map P. The first two convolution activation functions use the Relu function, and the last dimension reduction convolution uses a linear activation function instead of the Relu function. The input image is subjected to 1X1 convolution and batch normalization and matrix addition with the obtained feature map P, and the output feature map of a layer of linear bottleneck module is obtained through a linear activation function. In the present invention, three layers of linear bottleneck modules and four layers of linear bottleneck modules are cascaded, first increasing the dimension and then reducing the dimension, and a linear activation function is used to reduce the number of input image feature parameters, reduce the amount of calculation, and extract more image feature information, effectively preventing the feature from losing too much information in low-dimensional space and causing information collapse.

[0101] The bottleneck module of the axial attention, the attention calculation formula in this module is:

[0102]

[0103] Among them, y o represents the result of pooling the projected input at the random initial position 0 of the input; N is the number of position grids formed by the output of the preprocessing module; softmaxp represents the softmax function at the p = (a, b)th position; q represents the query, k is the key value, and v is the value; the learnable vector r represents the relative position encoding from p = (a, b) to 0 = (i, j); the superscript T represents the transposition; k p is the key value at position p, v p is the value at position p, is a learnable vector about the value from position p to point 0; is a learnable vector about the key value from position p to point 0; is the learnable vector for key query from position p to point 0; m is the number of columns in N.

[0104] The dual path converter module is as follows Figure 7 As shown;

[0105] The module has two built-in paths, one is the pixel path and the other is the global path. A cross-attention mechanism is used between the two paths (the input between the attention modules can come from the output on another path). The pixel features include a P2P attention module, a 1X1 convolution, two 3X3 convolutions, two batch normalization modules and a P2M attention module. The input pixel features are weighted by the P2P attention module, and the obtained pixel features are fused by a 1X1 convolution, a 3X3 convolution and a batch normalization module.

[0106] The fused features are input into a P2M attention module for calculation. The calculation formula of this module at pixel position a is:

[0107]

[0108] represents the output of attention at pixel position a, softmax n represents the softmax function of length N on the global path, represents a query at pixel position a, represents the key value at global position n, Represents the value at global position n.

[0109] The image features obtained after calculation are matrix-added with the initial values ​​of the pixel features and then enter the 3X3 convolution module and batch normalization module for feature fusion. The fused features are input into the Relu function and the pixel features are output.

[0110] The global path includes two fully connected layers, a batch normalization module, an M2P&M2M attention module, and a feature fusion network module. The input global memory is extracted through two fully connected layers, and the obtained features are input into an M2P&M2M attention module for calculation. The calculation formula of this module at position b is:

[0111]

[0112] represents the output of attention at the global position b, softmax n Indicates the length of the global path The softmax function, represents a query at global position b, Represents the key value of the pixel position at n and the global position matrix, v pmRepresents the value of the pixel position and the global position matrix at n.

[0113] The feature Z is obtained by matrix addition of the output of the M2P&M2M attention module and the initial value of the global memory. After being processed by the Relu activation function, the feature Z enters the feature fusion network module on the one hand, and performs a matrix addition operation with the output of the feature fusion network module on the other hand. Finally, it is processed by the Relu activation function to obtain the output mask.

[0114] The depthwise separable convolution module in the decoder is composed of a 3X3 depthwise convolution module and a 1X1 pointwise convolution module. Compared with the traditional convolutional neural network, the main advantage of using depthwise separable convolution is that it reduces the amount of calculation and the number of parameters, while improving the efficiency and speed of the model. Because in the depthwise convolution, each channel only needs to be convolved with one filter, instead of each channel needing to be convolved with all filters like the traditional convolution. Pointwise convolution further reduces the number of parameters by reducing the dimension of the input channel.

[0115] This structure of the improved MAX-DeepLab segmentation network in this embodiment can significantly reduce the size and computing requirements of the model while maintaining relatively high performance. Especially in resource-constrained environments such as mobile devices, it is easier to implement and deploy on different platforms, and can streamline the computing model, thereby achieving high-precision operations on smaller devices.

[0116] Specifically, the improved MAX-DeepLab segmentation network uses 32 TPU cores to train the model, with 100k iterations, 120 epochs, a batch size of 64, and uses the Radam and Lookahead optimization algorithms with a scheduled learning rate of 10 -3 , the backbone learning rate multiplier is 0.1, and the weight decay rate is 10 -4 , the descending path rate is 0.2. When the training model reaches convergence, that is, the model training gradient is close to 0.

[0117] The width measurement module is used to obtain the actual measured width and pixel scale of the mass concrete crack according to the predicted pixel width of the mass concrete crack using the conversion ratio between the pixel and the actual size, that is, the pixel scale α. The specific process is:

[0118] Place a ruler in the image with a ruler length of X. The pixel length of the ruler in the image is measured as L. The pixel scale α can be calculated by the following formula:

[0119] α=X / L

[0120] Among them, α represents the conversion ratio between pixels and actual size, that is, the pixel scale, X represents the actual length of the ruler, and L represents the pixel length of the ruler in the image.

[0121] It should be noted that the calculation of pixel scale requires that the length of the ruler in the image be accurately measured and that the ruler and the large-volume concrete crack defect be on the same plane. The width measurement module can be embedded at the end of the improved MAX-DeepLab segmentation network to directly obtain the actual width of the crack.

[0122] The crack defect image of the mass concrete area of ​​the tower to be inspected is pre-expanded and data enhanced using the improved StyleGAN-XL generative adversarial network. The obtained virtual data is then input into the trained improved Max-DeepLab segmentation network for quantitative analysis to obtain the width of the mass concrete cracks.

[0123] If it is identified as a large-volume concrete crack defect image, the measurement parameter threshold is set, the width of the large-volume concrete crack is calculated, and then the width is judged by the threshold. If it does not exceed the threshold range, the quality inspection is qualified, otherwise the quality inspection fails and an early warning is given and feedback is given.

[0124] Specifically, the width of the cracks in the mass concrete is checked. If the width of the cracks is less than 0.30 mm, the quality of the mass concrete is considered qualified. If the width of the cracks is greater than 0.30 mm, it is considered that an alarm check is required.

[0125] The present invention compares the improved STYLEGAN-XL generative adversarial network with the improved Max-DeepLab segmentation network on the segmentation effect of large-volume concrete cracks, and evaluates it on the generated 10,000 virtual data. When the generative adversarial network is not used for enhancement, the image segmentation effect of 500 images taken by the drone is extracted to a resolution of 512X512 as shown in Table 1, and the image segmentation effect when different generative adversarial networks are used for enhancement is shown in Table 2. The present invention adds the improved StyleGAN-XL generative adversarial network to the large-volume concrete crack segmentation network, solves the expansion and enhancement of high-pixel concrete images in deep learning, and quantitatively analyzes the cracks of large-volume concrete through the segmentation network, and then evaluates it with segmentation networks such as DETR, Axial-DeepLab, Panoptic-DeepLab, and Max-DeepLab-s on the validation set, and the test effect on the validation set has a more obvious improvement effect.

[0126] Table 1

[0127]

[0128] Table 2 Comparison of segmentation effects after processing by different enhancement methods

[0129]

[0130] Example 4

[0131] This embodiment provides a cable tower mass concrete image quality detection system, which is specifically used to expand and enhance the cable tower mass concrete image and complete its quality detection system. The workflow of the detection system is as follows: Figure 8 As shown, including:

[0132] The image collection and processing module uses a drone to take aerial photos of the cable tower, obtains multiple frames of images including different parts of the cable tower, and pre-processes the images;

[0133] A data annotation module is used to annotate the preprocessed images, annotate the large-volume concrete cracks in the preprocessed images, and distinguish their locations and ranges;

[0134] The data expansion and enhancement module is connected to the data annotation module, and uses the annotated images to train the improved StyleGAN-XL generative adversarial network to obtain a trained improved StyleGAN-XL generative adversarial network to generate sufficient number and high-quality virtual images;

[0135] The image segmentation module is connected with the data annotation module and the data expansion and enhancement module to obtain the annotated image and the generated virtual image, so as to train the improved Max-DeepLab segmentation network for detecting and quantifying the crack width of the massive concrete of the cable tower;

[0136] The early warning feedback module is connected to the image segmentation module. The image segmentation module detects the width of the cracks in the massive concrete of the cable tower, and promptly warns and feeds back the situation of the cracks in the massive concrete of the cable tower to the background system. The background system automatically issues an alarm and prompts relevant personnel to take measures.

[0137] The detection system of the present invention is highly customizable and can be used in a variety of different working environments, and can provide real-time feedback to the staff so that safety measures can be adjusted in a timely manner. This system has wide applicability and practicality and can be used for safety monitoring in a variety of working environments, including cable towers, skyscrapers, high-altitude exterior walls, and deep-sea pipelines. In practical applications, the present invention provides a new idea and solution to the problems of insufficient data acquisition and low data quality in various workplaces, and also provides the possibility of realizing the detection of large-volume concrete, further promoting the development of this field, which has great significance and broad application prospects, and also provides strong support and guarantee for the development of large-volume concrete quality detection technology.

[0138] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A high-quality method for generating a massive concrete image of a cable tower, characterized in that: The generation method includes the following contents: Use drones to take aerial photos of the cable tower, obtain images of the cable tower concrete, and mark the locations and ranges of cracks in the images; The improved StyleGAN-XL generative adversarial network is used to expand and enhance the obtained images to obtain high-quality large-volume concrete images of cable towers; The improved StyleGAN-XL generative adversarial network adopts a deep learning GAN network framework, including generator, discriminator and classifier modules: The generator is used to generate images of different resolutions, including an embedding module, a latent code, a mapping network module, a style code, a Fourier feature change, a dilated convolution module, a multi-layer retuned convolution module of different resolutions, and a TORGB module; the input image is preprocessed by the embedding module, and then enters the mapping network module together with the latent code, and the class label obtained by the classifier module is also input into the mapping network module to obtain the style code W, the style code enters the dilated convolution module after the Fourier feature change for feature extraction, and then enters the retuned convolution module of each layer for splicing, and then generates images of various resolutions through the TORGB module; if the resolution is less than 2048 2 When , it means that the resolution requirement is not met, and then it enters the discriminator; The discriminator includes two paths, the upsampling path is composed of upsampling, multi-head dilated attention convolution module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the downsampling path is composed of downsampling, mixed attention visual perception module, 1X1 convolution module, random 3X3 convolution layer, double linear upsampling, and four independent discriminators connected in sequence; the four independent discriminators of the two paths are respectively subjected to adversarial learning with the image features generated by the generator that do not meet the resolution requirements after being processed by the embedding module to generate high-quality images; Among the images that do not meet the resolution requirements, the relatively high-resolution images enter the upsampling path, and the relatively low-resolution images enter the downsampling path. After continuous confrontation between the two different paths of upsampling and downsampling, high-quality images are generated. The generated images that do not meet the resolution requirements but are of high quality are finally sent to the classifier module for the generator to continue to generate images that better meet the requirements; The hybrid attention visual perception module adopts a VIT module, and the multi-head dilated attention convolution module includes multiple attention heads, and a dilated convolution is set in each attention head, and the dilation rate of each dilated convolution is different; The retuning convolution module includes a modulation module, a demodulation module, a 3X3 convolution and a leaky Relu function, a double upsampling, a double tangent activation function, a double downsampling and a cropping image module. The generated style learned from the style code is affine transformed with the upper layer output feature, and then after being operated by the modulation module and the demodulation module, it enters the 3X3 convolution together with the upper layer output feature processed by the exponential moving average operation and is processed by the leaky Relu function. The image obtained after the double upsampling, the double tangent activation function and the double downsampling is cropped by the cropping image module to obtain the output feature map Q of the retuning convolution module; The TORGB module includes a modulation module and a 3X3 air-conditioning convolution, the number of holes is 2, and the feature map Q undergoes affine transformation and modulation module operation in the TORGB module, and enters the 3X3 air-conditioning convolution for processing together with the upper-layer output features processed by the exponential moving average operation.

2. The generation method according to claim 1, characterized in that: In the generator, the input image is first preprocessed by the embedding module, and the shallow features obtained are concatenated with a suitable latent code Z selected or interpolated from the pre-trained latent space and then enter the mapping network module for splicing. At the same time, the class label C obtained by the classifier module is concatenated by the mapping network module and mapped into a style code W. The style code W is concatenated with the Fourier feature change and then enters the dilated convolution module for feature re-extraction. After that, the 8×8 reconvolution module outputs an 8×8 image. At the same time, the 8×8 image is spliced ​​with W and then outputted through the 16×16 reconvolution module to generate a 16×16 image. At the same time, the 16×16 image is spliced ​​with W and then outputted into the 32×32 reconvolution module. After continuous iteration, the 1024×1024 image is spliced ​​with W and then outputted through the 2048×2048 reconvolution module to generate a 2048×2048 high-resolution image. Images of different resolutions are outputted through the TORGB module. If the resolution of the image is less than 2048, 2 Then enter the discriminator for operation.

3. The generation method according to claim 1, characterized in that: The acquired images are used to train the improved StyleGAN-XL generative adversarial network. During the training process, the number of network layers of the generator and the discriminator is gradually increased so that the network first learns large-scale structural information and then gradually shifts its attention to small-scale detail learning. At the beginning, 11 layers are used for progressive growth at a resolution of 16X16. Initially, 11 16×16 retuned convolution modules are set in series, and each retuned convolution module is a layer. Each time the resolution increases, 2 layers are cut off and 7 new layers are added. For the final stage of 2048X2048, only 5 layers are added without subtraction. There are 44 layers from the first 16×16 retuned convolution module to the last 2048×2048 retuned convolution module, and the maximum resolution is 2048X2048. A fixed growth plan is used to train each stage until the similarity index FID stops decreasing. After adding new layers, the existing layers with lower resolution will remain fixed and only the latest layers will be trained. The model is trained on V100 for 24 hours at a resolution of 2048X2048 and evaluated using the Fid evaluation indicator. Training is terminated when Fid is below 5. The trained improved StyleGAN-XL generative network is used to generate a large number of high-quality virtual images with a resolution of 2048X2048.

4. A method for detecting the image quality of a cable tower mass concrete, characterized in that: The detection method uses the generation method described in any one of claims 1 to 3 to obtain a virtual image, and uses a large number of virtual images as a target data set R of the segmentation network; The segmentation network adopts an improved MAX-DeepLab segmentation network, and the improved MAX-DeepLab segmentation network is trained using a target data set R for large-volume concrete crack width detection.

5. The detection method according to claim 4, characterized in that: The entire network of the improved Max-DeepLab segmentation network adopts a Transformer framework, including an encoder and a decoder; The encoder consists of a preprocessing module, a 3-layer linear bottleneck module, a 4-layer linear bottleneck module, a 6-layer axial attention bottleneck module, and a dual-path Transformer module; The input image is processed by the preprocessing module, and then passes through a 3-layer linear bottleneck module and a 4-layer linear bottleneck module to extract features to obtain a pixel feature map A. The pixel feature map A enters a 6-layer axial attention bottleneck module and then enters the pixel path of the dual-path Transformer module. The dual-path Transformer module also inputs a global memory, which is a trainable parameter. The decoder includes a plurality of stacked decoder bottleneck modules, each of which includes four 1X1 convolutions and a depth-separable convolution module. On the pixel path, each stacked decoder bottleneck module has two inputs, one is the feature output obtained in the previous layer, and the other is the pixel feature map A output by the 4-layer linear bottleneck module. The output of multiple stacked decoder bottleneck modules is the decoding feature E. After being processed by a 5X5 convolution module and a 1X1 convolution module, the decoding feature E is matrix multiplied with the mask features extracted by two fully connected layers on the global path to obtain a mask image; finally, after passing through a batch normalization module, the final mask image is output for quantitative measurement of cracks.

6. The detection method according to claim 5, characterized in that: The preprocessing module in the encoder includes a 7×7 standardized convolution, a group normalization, a ReLU activation function, and a pooling layer connected in sequence; The linear bottleneck module has three built-in 1X1 convolutions, one 3X3 convolution, and four batch normalization processing modules; the input feature map is sequentially processed by 1X1 convolution dimension increase and batch normalization, 3X3 depth convolution and batch normalization, 1X1 convolution dimension reduction and batch normalization to obtain a fused feature map P. The first two convolution activation functions use the Relu function, and the last dimension reduction convolution uses a linear activation function instead of the Relu function; the input image is again subjected to 1X1 convolution and batch normalization, and matrix addition is performed with the obtained feature map P, and after a linear activation function, an output feature map of a layer of linear bottleneck module is obtained.

7. The detection method according to claim 5, characterized in that: During the training of the improved MAX-DeepLab segmentation network, 32 TPU cores were used to train the model, with 100k iterations, 120 epochs, a batch size of 64, Radam and Lookahead optimization algorithms, and a scheduled learning rate of 10 -3 , the backbone learning rate multiplier is 0.1, and the weight decay rate is 10 -4 , the descent path rate is 0.2, and the training model is terminated when it reaches convergence.

8. A cable tower mass concrete image quality detection system, characterized in that: The detection system comprises: The image collection and processing module uses a drone to take aerial photos of the cable tower, obtains multiple frames of images including different parts of the cable tower, and pre-processes the images; A data annotation module is used to annotate the preprocessed images, annotate the large-volume concrete cracks in the preprocessed images, and distinguish their locations and ranges; The data expansion and enhancement module is connected to the data annotation module, and uses the annotated images to train the improved StyleGAN-XL generative adversarial network to obtain a trained improved StyleGAN-XL generative adversarial network to generate sufficient number and high-quality virtual images; The image segmentation module is connected with the data annotation module and the data expansion and enhancement module to obtain the annotated image and the generated virtual image, so as to train the improved Max-DeepLab segmentation network for detecting and quantifying the crack width of the massive concrete of the cable tower; The early warning feedback module is connected to the image segmentation module. The image segmentation module detects the width of the cracks in the massive concrete of the cable tower, and promptly warns and feeds back the situation of the cracks in the massive concrete of the cable tower to the background system. The background system automatically issues an alarm and prompts relevant personnel to take measures.

9. The detection system according to claim 8, characterized in that: When checking the width of cracks in mass concrete, if the width of the crack is less than 0.30 mm, the quality of the mass concrete is considered qualified. If the width of the crack is greater than 0.30 mm, an alarm check is considered necessary.

Citation Information

Cited By

  • Cable bent tower concrete crack identification method and system based on machine vision

    CN121725234A