A method for detecting concrete surface defects under defect-free sample conditions

Abstract

The present invention relates to a method for detecting concrete surface defects under defect-free sample conditions. The method comprises the following steps: constructing an image reconstruction model based on a VQ-VAE-2 network and training it using a normal concrete surface dataset to reconstruct a normal concrete surface image; geometrically transforming the image to be detected and inputting it into the image reconstruction model; calculating a block-level reconstruction error feature map between the transformed image to be detected and the output image; calculating a reconstruction error index for the image based on the block-level reconstruction error feature map to determine whether it is a concrete defect image, and generating a defect heat map representation for images determined to be concrete defects to achieve defect location. Compared with existing technologies, the present invention has the advantages of not requiring a large number of defect images and having high detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection, and in particular to a method for detecting concrete surface defects under defect-free sample conditions. Background Art

[0002] Concrete structures often suffer from various surface defects, such as cracks, weathering, spalling, and exposed rebar, which affect their service life and safety. Therefore, defect detection is a crucial task in the post-maintenance of concrete structures. However, traditional defect detection methods rely heavily on manual labor, resulting in low detection efficiency and susceptible to subjective experience, leading to false detections and missed detections. Traditional defect detection methods are increasingly unable to meet the current needs of concrete structure monitoring. With the rapid development of computer vision and deep learning technologies, intelligent object detection technology has been applied in many fields, providing new insights into concrete surface defect detection methods. However, deep learning models typically require a large number of defect samples for training, making the collection and annotation of concrete defects difficult, especially for concrete structures with a short service life. Therefore, developing efficient and accurate concrete surface defect detection methods based on computer vision and deep learning technologies, even in the absence of defect samples, has become an urgent need in the field of concrete surface defect detection.

[0003] Research on vision-based concrete surface defect detection methods often uses supervised convolutional neural networks, using models trained on defect samples to identify or segment defects in images. CN116109616A discloses a YOLOv5-based pavement crack detection and small facet fitting detection method, which uses an improved YOLOv5 network model to detect pavement concrete cracks and a VIT model to fit cracks to automatically identify cracks and calculate crack area. CN116246063A discloses a drone-based quantitative detection method for railway tunnel portal wall cracks, which uses an improved U-Net network to detect tunnel concrete cracks. To collect training samples, the patent uses drones to collect crack samples and uses a GAN model to expand the training set. These studies have shown that vision-based concrete surface defect detection methods have the advantages of high efficiency and high accuracy. However, most existing related studies rely on training sets, requiring the collection of a large number of defect images for model training. The lack of concrete defect image samples can lead to model overfitting, significantly reducing its accuracy and generalization ability. Currently, there is a lack of unsupervised methods for concrete surface defect detection. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for concrete surface defect detection under defect-free sample conditions. The model is trained using a defect-free sample data set, and defect detection is performed through reconstruction error. This solves the problem that the existing technology relies on a large number of defective samples to avoid model overfitting, and ensures high detection accuracy.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for detecting surface defects of concrete under defect-free sample conditions comprises the following steps:

[0007] Step 1) Build an image reconstruction model based on the VQ-VAE-2 network and train it using a normal concrete surface dataset to reconstruct normal concrete surface images;

[0008] Step 2) The image to be detected is geometrically transformed and then input into the image reconstruction model, and a block-level reconstruction error feature map of the geometrically transformed image to be detected and the output image is calculated;

[0009] Step 3) Calculate the image reconstruction error index based on the block-level reconstruction error feature map to determine whether it is a concrete defect image, and form a defect heat map representation for the image determined to be a concrete defect to achieve defect location.

[0010] The normal concrete surface dataset is obtained by using a camera to shoot a concrete structure surface without apparent defects, and preprocessing the captured image to divide it into multiple square images of preset sizes to form a normal concrete surface dataset.

[0011] During the training process of the image reconstruction model, the input image is continuously encoded by the lower-layer first encoder and the upper-layer second encoder. The upper layer obtains the upper-layer latent variable by vector quantization (VQ) of the output of the second encoder. After the upper-layer latent variable is decoded by the first decoder, it is combined with the output of the lower-layer first encoder and obtained by vector quantization to obtain the lower-layer latent variable. After upsampling, the upper-layer latent variable is input into the second decoder together with the lower-layer latent variable to obtain a reconstructed image.

[0012] The first encoder consists of 1 convolutional layer, 2 residual blocks and 1 convolutional layer in sequence, the second encoder consists of 2 residual blocks and 1 convolutional layer in sequence, the first decoder consists of 2 residual blocks, 1 convolutional layer and 1 deconvolutional layer in sequence, and the second decoder consists of 2 residual blocks, 1 convolutional layer and 2 deconvolutional layers of different sizes in sequence.

[0013] The loss function of the image reconstruction model is:

[0014]

[0015] Where: E 1( x ), E 2( x ) represent the output of the upper and lower encoders respectively, e 1. e 2 is the upper and lower layer latent variables after vector quantization, and is a hyperparameter, To stop calculating the gradient operator, The specific calculation formula is as follows:

[0016]

[0017] Where: It is t It is quantized into The number of encoder output values, yes The exponential moving average of , It is t In the iterations, it is transformed into The exponential sliding average of the encoder output value, γ is the weight parameter;

[0018] It is a hybrid function of MSE function, SSIM function and GMSD function:

[0019]

[0020] Among them, MSE function, SSIM function and GMSD function are evaluation indicators that characterize the similarity between the original image and the reconstructed image. x is the original image, To reconstruct the image.

[0021] The MSE function is used to evaluate the average difference between the reconstructed image and the original image at the pixel level. The lower the value, the more similar the two images are. The MSE function calculation formula is as follows:

[0022]

[0023] in, m 、 n are the number of pixel rows and columns of the image, mn represents the total number of image pixels, G ( i , j ) is the image G No. i Rankj The pixel value of the column, K ( i , j ) is the image K No. i Rank j The pixel value of the column.

[0024] The SSIM function compares three visual characteristics of the image: brightness, contrast and structure, among which brightness is evaluated by average grayscale, contrast is evaluated by standard deviation, and structure is evaluated by standardized image. That is, the SSIM function is a brightness contrast function. I ( x,y ), contrast contrast function c ( x,y ) and structure comparison function s ( x,y ) The combination of three functions is calculated as follows:

[0025]

[0026]

[0027]

[0028]

[0029] in, and They are images x and images y No. i The value of pixels; N is the total number of pixels in the image, C1, C2, and C3 are all constants to prevent the denominator from being 0; α, β, and γ are the weight parameters of the three functions in SSIM. .

[0030] The GMSD function is an image perception similarity evaluation index, which is determined based on the GMS feature map. The GMS feature map is determined based on the local position The transverse gradient amplitude and longitudinal gradient amplitude Sure:

[0031]

[0032] in, c is a constant to prevent the denominator from being 0,

[0033]

[0034] Horizontal gradient map r and longitudinal gradient map dUsing Prewitt operator 、 The convolution operation is performed on two images to calculate the Prewitt operator. The calculation formula is as follows:

[0035]

[0036] The GMS feature map reflects the similarity of the blocks, and the similarity index of the entire image is evaluated by the standard deviation of the GMS feature map, namely GMSD:

[0037]

[0038] in, for GMS ( i ) is the mean of .

[0039] The step 2) includes the following steps:

[0040] Step 21) Segment the image to be detected into a first preset number of blocks, disrupt the order of the blocks, and reassemble them into a new image as the input image of the model;

[0041] Step 22) Input the input image into the trained image reconstruction model and output the reconstructed image;

[0042] Step 23) The reconstructed image is divided into a first preset number of blocks, and the reconstruction error between the blocks of the reconstructed image and the corresponding blocks of the input image is calculated to generate a block-level reconstruction error feature map, wherein the reconstruction error is calculated using a hybrid function calculate.

[0043] The step 3) includes the following steps:

[0044] Step 31) Perform maximum pooling on the block-level reconstruction error feature map as an image-level reconstruction error index. Compare the reconstruction error index with a preset threshold. If the reconstruction error index is greater than the preset threshold, the image is judged to contain defects; otherwise, it is a normal image.

[0045] Step 32) For the concrete defect image determined to contain defects, the block-level reconstruction error feature map is restored to its original order and then upsampled. The upsampled feature map is color mapped to form a heat map representation of the defect to achieve defect location.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) The present invention utilizes the characteristic that the deep learning model depends on the training set, and uses easily available defect-free concrete images to train the deep learning model, forcing it to learn the ability to reconstruct defect-free concrete images, so that it will produce a large reconstruction error when reconstructing concrete defect images. Finally, the difference in reconstruction error between the defect-free concrete images and the concrete defect images is used to classify defect images. There is no need to rely on a large number of defect image data sets, and data acquisition is easy.

[0048] (2) The present invention adopts the VQ-VAE-2 network to realize image reconstruction, and uses the maximum pooling value of the block-level hybrid function reconstruction error feature map as the image reconstruction error indicator. In the detection stage, the image to be detected is subjected to random grid reorganization transformation, so that the reconstruction errors of the two types of images have significant differences, thereby improving the detection accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 This is a schematic diagram of data preprocessing results in an embodiment of the present invention;

[0051] Figure 3 This is a network architecture diagram of an image reconstruction model in an embodiment of the present invention;

[0052] Figure 4 Schematic diagram of the structure of the encoder 1 in an embodiment of the present invention;

[0053] Figure 5 Schematic diagram of the structure of the encoder 2 in an embodiment of the present invention;

[0054] Figure 6 Schematic diagram of the structure of the decoder 1 in an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the structure of a decoder 2 according to an embodiment of the present invention;

[0056] Figure 8 Schematic diagram of the loss curve in the model training phase;

[0057] Figure 9 Schematic diagram of geometric transformation based on random grid reorganization in an embodiment of the present invention;

[0058] Figure 10 An example of an input image to be detected and an image reconstructed by a model output in an embodiment of the present invention;

[0059] Figure 11 is a classification result diagram in an embodiment of the present invention;

[0060] Figure 12Schematic diagram of a thermal diagram of defects in an embodiment of the present invention;

[0061] Figure 13 This is an example of defect location results in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0063] This embodiment provides a method for detecting concrete surface defects under defect-free sample conditions, such as Figure 1 As shown, the following steps are included:

[0064] Step 1) Build an image reconstruction model based on the VQ-VAE-2 network and train it using a normal concrete surface dataset to achieve reconstruction of normal concrete surface images.

[0065] The normal concrete surface dataset is obtained by using a camera such as a drone, a camera, or a smart phone to photograph the surface of a concrete structure without apparent defects, and preprocessing the collected images to divide them into multiple square images of preset sizes, such as Figure 2 As shown, a normal concrete surface data set is formed.

[0066] The image reconstruction model in this embodiment is built based on the VQ-VAE-2 network, which is a variational autoencoder. Compared with traditional variational autoencoders, it has two major characteristics: discrete latent variables and hierarchical structure. The use of discrete latent variable distribution reduces the difficulty of learning the prior distribution on the one hand; on the other hand, the discretely distributed latent variables can approximate the continuous distribution, which is sufficient to meet the needs of the decoder to reconstruct the input image. In order to reconstruct a clear image, VQ-VAE-2 improves the encoder by adopting a hierarchical structure, and obtains a clearer output by fusing features of receptive fields of different sizes.

[0067] The image reconstruction model structure is as follows Figure 3 As shown in the figure, during the training process, the input image is continuously encoded by the lower-layer encoder 1 and the upper-layer encoder 2. The upper-layer output of the encoder 2 is vector quantized to obtain the upper-layer latent variable. After the upper-layer latent variable is decoded by the decoder 1, it is combined with the output of the lower-layer encoder 1 and vector quantized to obtain the lower-layer latent variable. After the upper-layer latent variable is upsampled, it is input into the decoder 2 together with the lower-layer latent variable to obtain the reconstructed image.

[0068] The model inputs an image of size 3×224×224 and outputs an image of size 3×224×224. The specific structures of encoder 1, encoder 2, decoder 1 and decoder 2 are as follows: Figure 4-Figure 7 shown.

[0069] Encoder 1 consists of a 64×112×112 convolutional layer, two residual blocks, and a 128×56×56 layer. Encoder 2 consists of two residual blocks and a 128×28×28 convolutional layer. Decoder 1 consists of two residual blocks, a 128×28×28 convolutional layer, and a 64×56×56 deconvolutional layer. Decoder 2 consists of two residual blocks, a 128×56×56 convolutional layer, a 64×112×112 deconvolutional layer, and a 3×224×224 deconvolutional layer.

[0070] The loss function of the image reconstruction model is divided into three terms, as shown below:

[0071]

[0072] Where: E 1( x ), E 2( x ) represent the output of the upper and lower encoders respectively, e 1. e 2 is the upper and lower layer latent variables after vector quantization, and is a hyperparameter, To stop calculating the gradient operator, The specific calculation formula is as follows:

[0073]

[0074] Where: It is t It is quantized into The number of encoder output values, yes The exponential moving average of , It is t In the iterations, it is transformed into The exponential sliding average of the encoder output value, γ is the weight parameter;

[0075] It is a hybrid function of MSE function, SSIM function and GMSD function:

[0076]

[0077] Among them, MSE function, SSIM function and GMSD function are evaluation indicators that characterize the similarity between the original image and the reconstructed image. x is the original image, To reconstruct the image.

[0078] The MSE function is easy to use with gradient descent algorithms because its curve is smooth, continuous, and can be differentiated everywhere. Furthermore, the gradient of the MSE function decreases as the error decreases, which facilitates rapid convergence. Therefore, the MSE function is widely used in deep learning. In anomaly detection, the MSE function is often used to evaluate the average pixel-level difference between the reconstructed image and the original image. The lower the value, the more similar the two images are. The MSE function calculation formula is as follows:

[0079]

[0080] in, m 、 n are the number of pixel rows and columns of the image, mn represents the total number of image pixels, G ( i , j ) is the image G No. i Rank j The pixel value of the column, K ( i , j ) is the image K No. i Rank j The pixel value of the column.

[0081] The SSIM function compares three visual characteristics of an image: luminance, contrast, and structure. Luminance is evaluated using the mean intensity; contrast is evaluated using the standard deviation; and structure is evaluated using a standardized image. The SSIM function is a luminance contrast function. I ( x, y ), contrast contrast function c ( x,y ) and structure comparison function s ( x,y ) The combination of three functions is calculated as follows:

[0082]

[0083]

[0084]

[0085]

[0086] in, and They are images x and images y No. i The value of pixels; N is the total number of pixels in the image, C1, C2, and C3 are all constants to prevent the denominator from being 0; α, β, and γ are the weight parameters of the three functions in SSIM. .

[0087] The GMSD function is an efficient image perception similarity evaluation indicator. Since image gradients are sensitive to changes in image structure, using only the gradient amplitude can achieve high accuracy. r and longitudinal gradient map d Usually the Prewitt operator is used 、 The convolution operation is performed on two images to calculate the Prewitt operator. The calculation formula of the Prewitt operator is as follows:

[0088]

[0089] In a local location The transverse gradient amplitude and longitudinal gradient amplitude Expressed as:

[0090]

[0091] On this basis, the GMS feature map can be calculated as follows:

[0092]

[0093] in, c is a constant to prevent the denominator from being zero.

[0094] The GMS feature map reflects the similarity of the blocks. The similarity index of the entire image is evaluated by the standard deviation of the GMS feature map, namely GMSD:

[0095]

[0096] in, for GMS ( i ) is the mean of .

[0097] This example uses the aforementioned image dataset and the image reconstruction model, trained on a computer equipped with an Intel(R) Core(TM) i7-10700K CPU and an NVIDIA GeForce RTX3070Ti GPU with 8GB of memory. Training was based on the following parameters: a learning rate of 0.0003 during the training phase; a batch size of 32; and a training iteration count of 160. The Adam optimizer was used to optimize the loss function. Figure 8 The changing curves of the total loss function and reconstruction error during training are shown.

[0098] Step 2) The image to be detected is geometrically transformed and then input into the image reconstruction model, and a block-level reconstruction error feature map of the geometrically transformed image to be detected and the output image is calculated.

[0099] Specifically, the following steps are included:

[0100] Step 21) Segment the image to be detected into k × k ( k (is any natural number, which can be selected according to the performance of the device) tiles, the order of the tiles is disrupted and recombined into a new image as the input image of the model;

[0101] Step 22) Input the input image into the trained image reconstruction model and output the reconstructed image;

[0102] Step 23) Segment the reconstructed image into k × k The reconstruction error between the reconstructed image block and the corresponding input image block is calculated to generate a block-level reconstruction error feature map, where the reconstruction error is obtained by using a hybrid function calculate.

[0103] This embodiment collected 5000 normal concrete pictures and 5000 crack pictures for detection. Figure 9 As shown in , the images to be tested are geometrically transformed, each image is divided into 4×4 image blocks, and then the order of the image blocks is disrupted to reconstruct new images. These transformed images are input into the trained image reconstruction model to obtain the output image. Examples of input and output are shown in Figure 10 The input and output of each image to be detected are divided into 4×4 image blocks, and the reconstruction error of the corresponding block is calculated to form a block-level reconstruction error feature map.

[0104] Step 3) Calculate the image reconstruction error index based on the block-level reconstruction error feature map to determine whether it is a concrete defect image, and form a defect heat map representation for the image determined to be a concrete defect to achieve defect location.

[0105] Specifically, the following steps are included:

[0106] Step 31) Perform maximum pooling on the block-level reconstruction error feature map as an image-level reconstruction error index. Compare the reconstruction error index with a preset threshold. If the reconstruction error index is greater than the preset threshold, the image is judged to contain defects; otherwise, it is a normal image.

[0107] Step 32) For concrete defect images that are judged to contain defects, restore their block-level reconstruction error feature maps to their original order and upsample them to a size of 224×224. Perform color mapping on the upsampled feature maps to form a heat map representation of the defects, thereby achieving defect location.

[0108] Use the detection method in step 3) to classify the input image and obtain Figure 11 The classification results are shown. This embodiment uses precision, recall, accuracy, and F1 score for evaluation, and the calculation formula is as follows:

[0109]

[0110]

[0111]

[0112]

[0113] The definitions of TP, TN, FP, and FN are shown in Table 1.

[0114] Table 1 Confusion matrix

[0115]

[0116] The final calculated results are shown in Table 2.

[0117] Table 2 Calculation results of precision, recall, accuracy, and F1 score

[0118]

[0119] For images identified as cracks, the block-level reconstruction error feature map is adjusted to the order of the original image, upsampled to 224×224, and then color mapped to generate a thermal map representation of the defect (such as Figure 12 As shown), the defect is located, and the positioning result is as follows Figure 13 shown.

[0120] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for detecting surface defects of concrete under defect-free sample conditions, characterized in that: The following steps are involved: Step 1) Build an image reconstruction model based on the VQ-VAE-2 network and train it using a normal concrete surface dataset to reconstruct normal concrete surface images; During the training process of the image reconstruction model, the input image is continuously encoded by the lower layer first encoder and the upper layer second encoder, the upper layer performs vector quantization on the output of the second encoder to obtain an upper layer latent variable, the upper layer latent variable is decoded by the first decoder, combined with the output of the lower layer first encoder and vector quantized to obtain a lower layer latent variable, the upper layer latent variable is upsampled and input into the second decoder together with the lower layer latent variable to obtain a reconstructed image; The loss function of the image reconstruction model is: ， Where: E 1( x ), E 2( x ) represent the output of the upper and lower encoders respectively, e 1. e 2 is the upper and lower layer latent variables after vector quantization, and is a hyperparameter, To stop calculating the gradient operator, The specific calculation formula is as follows: ， Where: It is t It is quantized into The number of encoder output values, yes The exponential moving average of , It is t In the iterations, it is transformed into The exponential sliding average of the encoder output value, γ is the weight parameter; It is a hybrid function of MSE function, SSIM function and GMSD function: ， Among them, MSE function, SSIM function and GMSD function are evaluation indicators that characterize the similarity between the original image and the reconstructed image. x is the original image, To reconstruct the image; Step 2) The image to be detected is geometrically transformed and then input into the image reconstruction model, and a block-level reconstruction error feature map of the geometrically transformed image to be detected and the output image is calculated; The step 2) includes the following steps: Step 21) Segment the image to be detected into a first preset number of blocks, disrupt the order of the blocks, and reassemble them into a new image as the input image of the model; Step 22) Input the input image into the trained image reconstruction model and output the reconstructed image; Step 23) The reconstructed image is divided into a first preset number of blocks, and the reconstruction error between the blocks of the reconstructed image and the corresponding blocks of the input image is calculated to generate a block-level reconstruction error feature map, wherein the reconstruction error is calculated using a hybrid function calculate; Step 3) Calculate the image reconstruction error index based on the block-level reconstruction error feature map to determine whether it is a concrete defect image, and form a defect heat map representation for the image determined to be a concrete defect to achieve defect location.

2. A method for detecting surface defects of concrete under defect-free sample conditions according to claim 1, characterized in that: The normal concrete surface dataset is obtained by using a camera to shoot a concrete structure surface without apparent defects, and preprocessing the captured image to divide it into multiple square images of preset sizes to form a normal concrete surface dataset.

3. The method for detecting surface defects of concrete under defect-free sample conditions according to claim 1, characterized in that: The first encoder consists of 1 convolutional layer, 2 residual blocks and 1 convolutional layer in sequence, the second encoder consists of 2 residual blocks and 1 convolutional layer in sequence, the first decoder consists of 2 residual blocks, 1 convolutional layer and 1 deconvolutional layer in sequence, and the second decoder consists of 2 residual blocks, 1 convolutional layer and 2 deconvolutional layers of different sizes in sequence.

4. The method for detecting surface defects of concrete under defect-free sample conditions according to claim 1, characterized in that: The MSE function is used to evaluate the average difference between the reconstructed image and the original image at the pixel level. The lower the value, the more similar the two images are. The MSE function calculation formula is as follows: ， in, m 、 n are the number of pixel rows and columns of the image, mn represents the total number of image pixels, G ( i , j ) is the image G No. i Rank j The pixel value of the column, K ( i , j ) is the image K No. i Rank j The pixel value of the column.

5. The method for detecting concrete surface defects under defect-free sample conditions according to claim 1, characterized in that: The SSIM function compares three visual characteristics of the image: brightness, contrast and structure, among which brightness is evaluated by average grayscale, contrast is evaluated by standard deviation, and structure is evaluated by standardized image. That is, the SSIM function is a brightness contrast function. I ( x,y ), contrast contrast function c ( x,y ) and structure comparison function s ( x,y ) The combination of three functions is calculated as follows: ， ， ， ， in, and They are images x and images y No. i The value of pixels; N is the total number of pixels in the image, C1, C2, and C3 are all constants to prevent the denominator from being 0; α, β, and γ are the weight parameters of the three functions in SSIM. .

6. The method for detecting concrete surface defects under defect-free sample conditions according to claim 1, characterized in that: The GMSD function is an image perception similarity evaluation index, which is determined based on the GMS feature map. The GMS feature map is determined based on the local position The transverse gradient amplitude and longitudinal gradient amplitude Sure: ， in, c is a constant to prevent the denominator from being 0, ， Horizontal gradient map r and longitudinal gradient map d Using Prewitt operator 、 The convolution operation is performed on two images to calculate the Prewitt operator. The calculation formula is as follows: ， The GMS feature map reflects the similarity of the blocks, and the similarity index of the entire image is evaluated by the standard deviation of the GMS feature map, namely GMSD: ， in, for GMS ( i ) is the mean of .

7. The method for detecting surface defects of concrete under defect-free sample conditions according to claim 1, characterized in that: The step 3) includes the following steps: Step 31) Perform maximum pooling on the block-level reconstruction error feature map as an image-level reconstruction error index. Compare the reconstruction error index with a preset threshold. If the reconstruction error index is greater than the preset threshold, the image is judged to contain defects; otherwise, it is a normal image. Step 32) For the concrete defect image determined to contain defects, the block-level reconstruction error feature map is restored to its original order and then upsampled. The upsampled feature map is color mapped to form a heat map representation of the defect to achieve defect location.

Images