QUICK AND INTELLIGENT IR7-EC NETWORK BASED CLASSIFICATION METHOD FOR CONCRETE IMAGE CRACK TYPE

Abstract

An intelligent method for efficiently classifying concrete cracks from large amounts of image data is proposed, named inverted residual (IR) 7-Efficient Channel Attention and Convolutional Block Attention Module (EC) network. The IR7-EC network consists of a convolutional layer, seven inverted residual-ECA structures, a CBAM attention mechanism, a pooling layer, and multiple fully connected layers that are sequentially connected. The inverted residual-ECA structure consists of two components: a depthwise separable convolution-based inverted residual structure and an ECA attention mechanism. The new inverted residual structure facilitates the feature extraction of concrete cracks. Compared to conventional network structures like VGG and Resnet, the proposed IR7-EC network excels in both accuracy and efficiency. Once the IR7-EC network is fully trained, it can accurately classify various types of concrete cracks in captured images. This method offers several advantages, including a small number of network parameters, fast training convergence speed, and precise classification results.

Description

CROSS REFERENCE TO THE RELATED APPLICATIONS

This application is based upon and claims priority to Chinese Patent Application No. 202210857092.X, filed on Jul. 20, 2022, the entire contents of which are incorporated herein by reference.

TECHNICAL FIELD

The present invention pertains to the technical field of classifying types of concrete cracks based on inverted residual (IR) 7-ECA and CBAM (EC) networks. Specifically, it focuses on a fast and intelligent classification method for vision-based concrete crack type classification.

BACKGROUND

Concrete structures such as bridges, dams, and other civil infrastructures are susceptible to degradation and damage during operation. This can have a significant impact on safety, making it crucial to diagnose potential damage. The advancement of unmanned aerial vehicle and wireless transmission technologies has enabled the efficient collection of mass data on concrete structures. This serves as a basis for the development of intelligent classification. Compared to traditional methods for classifying structural damage, vision-based deep learning technology offers several advantages. It is highly efficient, precise, and objective when it comes to classifying cracks.

Deep learning neural network is widely used to solve multi-classification problems in the computer field. Its purpose was to create a universal and efficient network that could be applied in various scenarios with high precision. However, when this type of network is used for classifying concrete cracks, it only needs to classify up to ten different types of cracks. As a result, using the universal network can lead to parameter redundancy and waste of training time and hardware memory. Therefore, it is necessary to develop a special neural network for classifying concrete cracks, with light weight, high network convergence speed, and high classification precision.

SUMMARY

In order to solve the above problems, the present invention provides a crack type classification method based on an inverted residual (IR) 7-ECA and CBAM (EC) network model special for mass concrete image crack classification.

A quick and intelligent IR7-EC network based classification method for a concrete image crack type includes the following steps:

• • building an IR7-EC network model, where the IR7-EC network model includes a convolutional layer, 7 inverted residual-efficient channel attention (ECA) structures, a convolutional block attention module (CBAM) attention mechanism, a pooling layer and a plurality of fully connected layers which are connected sequentially, and
  • the inverted residual-ECA structure is formed by connecting a depthwise separatable convolution based inverted residual structure and an ECA attention mechanism;
  • inputting an actually-shot concrete image to be detected into a trained IR7-EC network model, and obtaining an initial feature through the convolutional layer;
  • making the initial feature sequentially pass the 7 inverted residual-ECA structures, so as to obtain an enhanced concrete crack feature extraction map;
  • inputting the enhanced concrete crack feature extraction map into the CBAM attention mechanism, and conducting feature extraction at a channel and space level, so as to obtain a feature extraction map containing more crack information; and
  • conducting data dimension reduction and two-dimensional to one-dimensional transformation on the feature extraction map, and classifying an input concrete crack image according to extracted one-dimensional crack image feature information, so as to output an classification result.

Preferably, the building an IR7-EC network model specifically includes the following step:

• • establishing 13 layers of network frameworks connected sequentially, where
  • a 1st layer includes a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer, and a hardswish activation function, a step size being 2, the number of input channels being 3, and the number of output channels being 16;
  • a 2nd layer to a 8th layer are the inverted residual-ECA structures, and each layer specifically includes four parts:
  • a first part including: a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and an ReLU6 activation function;
  • a second part including: a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer, and an ReLU6 activation function;
  • a third part being: an ECA attention mechanism, a structure of which includes an average pooling layer, a convolutional layer with a convolutional kernel of 1×1, and a sigmoid function; and
  • a fourth part including: a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and a linear activation function;
  • a 9th layer includes a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer and a hardswish activation function, a step size being 1, the number of input channels being 96, and the number of output channels being 576;
  • a 10th layer is the CBAM attention mechanism and specifically includes two parts:
  • a first part being: a channel attention mechanism, which includes an average pooling layer, a maximum pooling layer, the fully connected layer 1, an ReLU6 activation function, the fully connected layer 2, and a sigmoid function; and
  • a second part being: a space attention mechanism, which includes an average pooling layer, a maximum pooling layer, a convolutional layer with a convolutional kernel of 7×7, and a sigmoid function;
  • a 11th layer is an average pooling layer;
  • a 12th layer is a one-dimensional convolutional layer, a hardswish activation function, and a dropout layer with a sparsity ratio of 0.2; and
  • a 13th is a one-dimensional convolutional layer.

Preferably, training of the IR7-EC network model includes the following steps:

• • collecting crack images, establishing an image data set of different crack types, and conducting pre-processing;
  • inputting a pre-processed image data set into the 1st layer, and conducting preliminary feature extraction, data normalization and nonlinear transformation on image sample data sequentially through the convolutional layer, the batch normalization layer and the hardswish activation function;
  • inputting a feature preliminarily extracted through the 1st layer into the 2nd layer to the 8th layer, making the feature sequentially pass the 7 inverted residual-ECA structures, inputting the feature into the first part of the inverted residual-ECA structure for processing so as to augment a dimension of an input data channel, inputting data after channel dimension augmentation into the second part so as to conduct deep feature extraction, inputting data after deep feature extraction into the third part, highlighting some channels with higher network precision by weighting network channels through the ECA attention mechanism, so as to obtain the enhanced concrete crack feature extraction map, inputting the enhanced concrete crack feature extraction map into the fourth part so as to reduce a dimension of a data channel, and making the data pass all the inverted residual-ECA structures sequentially according to similar operations;
  • inputting the data obtained through the 2nd layer to the 8th layer into the 9th layer for feature extraction;
  • inputting data after feature extraction through the 9th layer into the CBAM attention mechanism of the 10th layer, and further conducting feature extraction at a channel and space level on the data, so as to obtain a feature extraction map containing more crack information;
  • inputting the feature extraction map obtained through the 10th layer into the 11th layer, the 12th layer and the 13th layer sequentially so as to conduct data dimension reduction and two-dimensional to one-dimensional transformation, and classifying an input concrete crack image according to extracted one-dimensional crack image feature information, so as to output an classification result; and
  • comparing the classification result obtained through the 12th layer and a true result, computing an error through a loss function, and updating a network parameter through an optimization function until a network is fitted.

Preferably, normalization of the batch normalization layer specifically includes:

${\mu }_{ℬ}=\frac{1}{m}\sum _{i=1}^m{x}_i$ ${\sigma }_{ℬ}^2=\frac{1}{m}\sum _{i=1}^m{\left(x_i-{\mu }_{ℬ}\right)}^2$ ${\hat{x}}_i=\frac{x_i-{\mu }_{ℬ}}{\sqrt{{\sigma }_{ℬ}^2+ϵ}}$ $y_i\leftarrow \gamma {\hat{x}}_i+\beta$

where x_i is a feature map before input into batch normalization, y_i is a feature map after output from batch normalization, m is the number of feature maps input into the layer in a current training batch, and γ and β are variables that change with update of a network gradient;

• • data passing the ReLU6 activation function is processed nonlinearly through the following formula:

f(x_i)=min(max(x_i,0),6)

• • where x_i is a feature map before input into the ReLU6 activation function, and f(x_i) is a feature map after output from the ReLU6 activation function; and
  • data passing the hardswish activation function is processed nonlinearly through the following formula:

$\mathrm{Hardswish}\left(x\right)=\{\begin{array}{cc}0& \mathrm{if}x\le -3\\ x& \mathrm{if}x\ge +3\\ x·\left(x+3\right)/6& \mathrm{otherwise}\end{array}$

• • where x is a feature map before input into the hardswish activation function, and f(x) is a feature map after output from the hardswish activation function.

Preferably, the highlighting some channels with higher network precision by weighting network channels through the ECA attention mechanism, so as to obtain the enhanced concrete crack feature extraction map includes the following step:

• • conducting cross-channel interaction on data passing the ECA attention mechanism in each layer, so as to obtain the enhanced concrete crack feature extraction map, by using the following formula:

$k=\psi \left(C\right)={❘\frac{{\log }_2\left(C\right)}{\gamma }+\frac{b}{\gamma }❘}_{\mathrm{odd}}{E}_s\left(F\right)=\sigma \left(f^{k*k}\left[\mathrm{AvgPool}\left(F\right)\right]\right)$

where |t|_odd represents a nearest odd t; C is the number of channels of data input into the ECA attention mechanism, and γ and B are two hyper-parameters, γ being 2, and b being 1; and E_s(F) is the ECA attention mechanism, σ is a sigmoid operation, f^k*k[⋅] represents a convolutional operation of k*k, F is an input feature map, and AvgPool( ) is average pooling.

Preferably, the inputting data after feature extraction through the 9th layer into the CBAM attention mechanism of the 10th layer, and further conducting feature extraction at a channel and space level on the data, so as to obtain a feature extraction map containing more crack information includes the following steps:

• • using average pooling and maximum pooling to aggregate space information of feature mapping, compressing a space dimension of an input feature map, and conducting summation and merging element by element, so as to generate a channel attention map, through the following formula:

M _c(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

• • where M_c represents channel attention, MLP( ) is composed of the fully connected layer 1, the ReLU6 activation function, and the fully connected layer 2, σ is a sigmoid operation, F is an input feature map, AvgPool( ) is average pooling, MaxPool( ) is maximum pooling, and M_s represents the space attention mechanism; and
  • using average pooling and maximum pooling to compress the input feature map in a space attention module, so as to obtain the feature extraction map containing more crack information through the following formula:

M _s(F)=σ(f^7*7[AvgPool(F),MaxPool(F)])

• • where M_s represents the space attention mechanism, σ is a sigmoid operation, f^7*7[⋅] represents a convolutional operation of 7*7, F is an input feature map, AvgPool( ) is average pooling, and MaxPool( ) is maximum pooling.

Preferably, the method further includes:

• • conducting sparsification on data passing the dropout layer in each layer, so as to avoid network over-fitting through the following formulas:

r _j ^(l)˜Bernoulli(p)

{tilde over (y)} ^(l) =r ^(l) *y ^(l)

• • where a Bernoulli(p) function is used to generate a vector of a probability r_j^(l), such that a neuron stops working with probability p, y^(l) is an upper layer of an output feature map, and {tilde over (y)}^(l) is a feature map output after passing the dropout layer.

Preferably, the computing an error through a loss function includes the following step:

• • computing loss of a network through the following formula:

$\mathrm{Loss}\left(y_{o,c},p_{o,c}\right)=-\sum _{c=1}^M{y}_{o,c}\log \left(p_{o,c}\right)$

• • where Loss(y_o,c, p_o,c) is a loss function between a network predicted value and a true value, M is the number of classes, 7 and y_o,c are true values, and p_o,c is the network predicted value.

Preferably, the method further includes:

• • optimizing internal parameters of a network through the following formulas:

f(θ)=Loss(y_o,c,p_o,c)

g _t=∇_θf_t(θ_t−1)

m _t=β₁·m_t−1+(1−β₁)·g_t

v _t=β₂·v_t−1+(1−β₂)·g_t²

{circumflex over (m)}t=m _t/(1−β₁^t)

{circumflex over (v)} _t =v _t|(1−β₂^t)

θ_y=θ_t−1−α·{circumflex over (m)}_t/(√{square root over ({circumflex over (v)}_t)}+ϵ)

where Loss(y_o,c, p_o,c) is a loss function between a network predicted value and a true value, θ is a parameter to be updated in a model, g_t is a gradient obtained by conducting derivation on θ of the loss function f(θ), β₁ is a first-order moment attenuation coefficient, β₂ is a second-order moment attenuation coefficient, m_t is an expectation of the gradient g_t, v_t is an expectation of g_t², {circumflex over (m)}_t is an offset correction of m_t, {circumflex over (v)}_t is an offset correction of v_t, θ_t□1 is a parameter before network update, and θ_t is a parameter after network update.

The present invention has the beneficial effects:

According to the present invention, based on the inverted residual structure and other machine vision algorithms, the IR7-EC network model special for mass concrete image crack classification is created. Compared with a current popular network of computer vision, the IR7-EC network model has a smaller number of parameters and shorter training time, and maintains higher concrete crack classification precision.

The present invention avoids disadvantages such as parameter redundancy, long training time and large hardware memory occupation of current general classification networks of alexnet, vgg16, resnet50, Google net and mobilenet_v3_large, has characteristics and advantages such as a small number of network model parameters, a high training convergence speed and high concrete crack classification precision, forms a special model for intelligently and efficiently classifying cracks from concrete image big data, and has great engineering application potential.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow diagram of using an inverted residual (IR) 7-ECA and CBAM (EC) network to classify a concrete crack according to Example 1 of the present invention;

FIG. 2 shows schematic diagrams of 7 types of concrete crack images according to Example 1 of the present invention;

FIG. 3 shows an evaluation criterion of crack types of a transverse crack, a vertical crack and an oblique crack of concrete according to Example 1 of the present invention;

FIGS. 4A and 4B are comparison diagrams of a traditional inverted residual structure and a combination of an inverted residual structure and an efficient channel attention (ECA) attention mechanism according to Example 1 of the present invention;

FIG. 5 is a schematic structural diagram of an IR7-EC network according to Example 1 of the present invention;

FIG. 6 shows visual maps of crack results identified by using an IR7-EC network according to Example 1 of the present invention;

FIGS. 7A-7D show comparison diagrams of training results between an IR7-EC model and different convolutional neural network (CNN) models; and

FIGS. 8A-8F show comparison diagrams of confusion matrices of test set results between an IR7-EC model and different CNN models.

DETAILED DESCRIPTION OF THE EMBODIMENTS

For making objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in further detail below in conjunction with the accompanying drawings and examples. It should be understood that specific examples described herein are merely used to explain the present invention, and are not used to limit the present invention.

Example 1

The present invention provides a quick and intelligent inverted residual (IR) 7-ECA and CBAM (EC) network based classification method for a concrete image crack type. A flow diagram of using an IR7-EC network to classify concrete crack is as shown in FIG. 1.

TABLE 1
Computer platform and environment configuration
implemented in the example
Hardware and
software platform Type parameter
Operation system  Windows 10
CPU               Intel(R) Xeon(R) Gold 5222
                  CPU @ 3.80 GHz 3.79 GHz
GPU               NVIDIA Quadro P2200
Internal storage  64.0 GB
Programming       Anaconda3
environment       CUDA10.2
                  Python3.6
                  pytorch

With reference to FIG. 1, the present invention provides an IR7-EC network model special for mass concrete image crack classification, which specifically includes the following steps:

• • step 1, crack images are collected, and a crack type classification image data set is established, where the crack type classification image data set includes a crack type classification image training set and validation set;
  • step 2, an image to be input into a neural network model is pre-processed;
  • step 3, the neural network model is built by fusing a machine vision algorithm on the basis of an inverted residual structure, image sample data is input into the neural network model so as to train the neural network model until a neural network reaches a converged state, and the network model is saved; and
  • step 4, a concrete crack image to be detected is input into a trained neural network, so as to obtain a type of a concrete crack in a concrete crack image.

Specifically, in step 1, after crack images are collected, the concrete crack images are manually classified, which include: images of a transverse crack, a vertical crack, an oblique crack, a mesh crack, an irregular crack, a hole, and a background, as shown in FIG. 2. Hole-augmentation is to manually augment a hole concrete crack data set on the basis of a digital image processing technology (random pixel addition, color temperature change, perspective transformation, horizontal flipping, random pixel zeroing, motion blur, gaussian noise, unequal scaling or other technologies are randomly combined to act on an original hole crack map), and then augmented data is merged into an original hole crack class data set, so as to solve a problem of scarce data. An evaluation criterion of crack types of a transverse crack, a vertical crack and an oblique crack of concrete is as shown in FIG. 3. The training set and the validation set include class labels corresponding to the above 7 types of images and image types, which include 10,000 images in total, the number of each type of images is basically the same, and a ratio of the number of images in the training set to the validation set is 7:3.

Specifically, in step 2, pre-processing of the training set includes random image horizontal flipping and image normalization, where an average value required for normalization is set as [0.485, 0.456, 0.406], and a variance is set as [0.229, 0.224, 0.225]. Pre-processing of the validation set includes normalization, where parameters are set as [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225].

In the example, the step of building a neural network model in step 3 includes the following step:

• • 13 layers of network frameworks connected sequentially are established, as shown in FIG. 5. Specific internal parameters are as shown in Table 2, where conv2d represents a convolutional layer, pool represents a pooling layer, NBN represents no batch normalization layer, HS represents a hardswish activation function, RE represents a ReLU6 activation function, and k represents the number of crack types.

TABLE 2
IR7-EC network structure and specific internal parameters
Input image			The	Whether to
(Length, width,	Operation name,		number of	use an ECA	Nonlinear
the number of	convolutional	Expansion	output	attention	activation	Step
channels)	#Z,899, size	multiple	channels	mechanism	function	size
224 × 224 × 3	conv2d		16		HS	2
112 × 112 × 16	Inverted Residual-BCA, 3 × 3	16	16		RE	2
56 × 56 × 16	Inverted Residual-BCA, 3 × 3	72	24		RE	2
28 × 28 × 24	Invented Residual-BCA, 5 × 5	88	24	✓	RE	1
28 × 28 × 24	Inverted Rasidual-BCA, 3 × 3	96	40		HS	2
14 × 14 × 46	Inverted Residual-BCA, 3 × 3	240	48	✓	HS	1
14 × 14 × 48	Inverted Residual-BCA, 5 × 5	288	96	✓	HS	2
7 × 7 × 96	Inverted Residual-BCA, 5 × 5	576	96	✓	HS	1
7 × 7 × 9	conv2d, 1 × 1		576	✓	HS	1
Channel Attention
Spatial Attention
7 × 7 × 576	, 7 × 7					1
1 × 1 × 576	conv2d, 1 × 1, NBN.dropout		512		HS	1
1 × 1 × 1024	conv2d, 1 × 1, NBN		k			1
indicates data missing or illegible when filed

A 1st layer with a convolutional kernel of 3×3, a batch normalization layer, and a hardswish activation function, a step size being 2, the number of input channels being 3, and the number of output channels being 16.

A 2nd layer are the inverted residual-efficient channel attention (ECA) structures. Comparison between the inverted residual-ECA structure and a traditional inverted residual structure is as shown in FIGS. 4A and 4B. The inverted residual-ECA structure specifically includes four parts: a first part including a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and an ReLU6 activation function; a second part including a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer, and an ReLU6 activation function; a third part being an ECA attention mechanism, a structure of which includes an average pooling layer, a convolutional layer with a convolutional kernel of 1×1, and a sigmoid function; and a fourth part structurally including a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and a linear activation function.

A 9th layer includes a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer and a hardswish activation function, a step size being 1, the number of input channels being 96, and the number of output channels being 576.

A 10th layer is the CBAM attention mechanism and specifically includes two parts: a first part being a channel attention mechanism, which includes an average pooling layer, a maximum pooling layer, the fully connected layer 1, an ReLU6 activation function, the fully connected layer 2, and a sigmoid function; and a second part being a space attention mechanism, which includes an average pooling layer, a maximum pooling layer, a convolutional layer with a convolutional kernel of 7×7, and a sigmoid function.

A 11th layer is an average pooling layer.

A 12th layer is a one-dimensional convolutional layer, a hardswish activation function, and a dropout layer with a sparsity ratio of 0.2.

A 13th is a one-dimensional convolutional layer.

With reference to FIG. 1, the step that image sample data is input into a neural network model so as to train the neural network model includes the following steps:

• • pre-processed image sample data is input into the 1st layer, and preliminary feature extraction, data normalization and nonlinear transformation are conducted on the image sample data sequentially through the convolutional layer, the batch normalization layer and the hardswish activation function;
  • a feature preliminarily extracted through the 1st layer is input into the 2nd layer to the 8th layer, the feature sequentially passes the 7 inverted residual-ECA structures, data is input into the first part of the inverted residual-ECA structure for processing so as to augment a dimension of an input data channel, an augmentation multiple being the expansion multiple as shown in Table 1, data after channel dimension augmentation is input into the second part so as to conduct deep feature extraction, data after deep feature extraction is input into the third part, some channels with higher network precision are highlighted by weighting network channels through the ECA attention mechanism, so as to obtain an enhanced concrete crack feature extraction map, the enhanced concrete crack feature extraction map is input into the fourth part so as to reduce a dimension of a data channel, the number of channels may decrease to the number of output channels in Table 1, when the number of input channels is equal to the number of output channels and the step size is 1, the inverted residual-ECA structure may be connected in a shortcut manner, and the data may pass all the inverted residual-ECA structures sequentially according to similar operations;
  • the data obtained through the 2nd layer to the 8th layer is input into the 9th layer for feature extraction;
  • data after feature extraction through the 9th layer is input into the CBAM attention mechanism of the 10th layer, and feature extraction at a channel and space level is further conducted on the data, so as to obtain a feature extraction map containing more crack information;
  • the feature extraction map containing more crack information obtained through the 10th layer is input into the 11th layer, the 12th layer and the 13th layer sequentially so as to conduct data dimension reduction and two-dimensional to one-dimensional transformation, and an input concrete crack image is identified according to extracted one-dimensional crack image feature information, so as to output an classification result; and
  • the classification result obtained through the 12th layer is compared with a true result, an error is computed through a loss function, and a network parameter is updated through an optimization function until a network is fitted.

Specifically, in order to prevent a vanishing network gradient, normalization of the batch normalization layer in each layer is conducted through the following formulas:

${\mu }_{ℬ}=\frac{1}{m}\sum _{i=1}^m{x}_i{\sigma }_{ℬ}^2=\frac{1}{m}\sum _{i=1}^m{\left(x_i-{\mu }_{ℬ}\right)}^2{\hat{x}}_i=\frac{x_i-{\mu }_{ℬ}}{\sqrt{{\sigma }_{ℬ}^2+ϵ}}{y}_i\leftarrow \gamma {\hat{x}}_i+\beta$

where x_i is a feature map before input into batch normalization, y_i is a feature map after output from batch normalization, m is the number of feature maps input into the layer in a current training batch, and γ and β are variables that change with update of a network gradient.

Specifically, data passing the ReLU6 activation function in each layer is processed nonlinearly through the following formula:

f(x_i)=min(max(x_i,0),6)

• • where x_i is a feature map before input into the ReLU6 activation function, and f(x_i) is a feature map after output from the ReLU6 activation function.

Specifically, data passing the hardswish activation function in each layer is processed nonlinearly through the following formula:

$\mathrm{Hardswish}\left(x\right)=\{\begin{array}{cc}0& \mathrm{if}x\le -3\\ x& \mathrm{if}x\ge +3\\ x·\left(x+3\right)/6& \mathrm{otherwise}\end{array}$

• • where x is a feature map before input into the hardswish activation function, and f(x) is a feature map after output from the hardswish activation function.

Specifically, cross-channel interaction is conducted on data passing the ECA attention mechanism in each layer, so as to obtain the enhanced concrete crack feature extraction map, by using the following formula:

$k=\psi \left(C\right)={❘\frac{{\log }_2\left(C\right)}{\gamma }+\frac{b}{\gamma }❘}_{\mathrm{odd}},E_s\left(F\right)=\sigma \left(f^{k*k}\left[\mathrm{AvgPool}\left(F\right)\right]\right)$

• • where |t|_odd represents a nearest odd t; C is the number of channels of data input into the ECA attention mechanism, and γ and B are two hyper-parameters, γ being 2, and b being 1; and E_s(F) is the ECA attention mechanism, σ is a sigmoid operation, f^k*k[.] represents a convolutional operation of k*k, F is an input feature map, and AvgPool( ) is average pooling.

Specifically, average pooling and maximum pooling are used to aggregate space information of feature mapping, a space dimension of an input feature map is compressed, and summation and merging are conducted element by element, so as to generate a channel attention map, through the following formula.

M _c(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

• • where M_c represents channel attention, MLP( ) is composed of the fully connected layer 1, the ReLU6 activation function, and the fully connected layer 2, σ is a sigmoid operation, F is an input feature map, AvgPool( ) is average pooling, MaxPool( ) is maximum pooling, and M_s represents the space attention mechanism.

Specifically, average pooling and maximum pooling are used to compress the input feature map in a space attention module, so as to obtain the feature extraction map containing more crack information through the following formula.

M _s(F)=a(f^7*7[AvgPool(F),MaxPool(F)])

• • where M_s represents the space attention mechanism, σ is a sigmoid operation, f^7*7[ ] represents a convolutional operation of 7*7, F is an input feature map, AvgPool( ) is average pooling, and MaxPool( ) is maximum pooling.

Specifically, sparsification is conducted on data passing the dropout layer in each layer, so as to avoid network over-fitting through the following formulas:

r _j ^(l)˜Bernoulli(p)

y ^(l) =r ^(l) *y ^(l)

where a Bernoulli(p) function is used to generate a vector of a probability r_j^(l), such that a neuron stops working with probability p, y^(l) is an upper layer of an output feature map, and {tilde over (y)}^(l) is a feature map output after passing the dropout layer.

Specifically, loss of a network is computed through the following formula:

$\mathrm{Loss}\left(y_{o,c},p_{o,c}\right)=-\sum _{c=1}^M{y}_{o,c}\log \left(p_{o,c}\right)$

• • where Loss(y_o,c, p_o,c) is a loss function between a network predicted value and a true value, M is the number of classes, 7 and y_o,c are true values, and p_o,c is the network predicted value.

Specifically, internal parameters of a network are optimized through the following formulas:

f(θ)=Loss(y_o,c,p_o,c)

g _t=∇_θf_t(θ_t−1)

m _t=β₁·m_t−1+(1−β₁)·g_t

v _t=β₂·v_t−1+(1−β₂)·g_t²

{circumflex over (m)} _t =m _t/(1−β₁^t)

{circumflex over (v)} _t =v _t/(1−β₂^t)

θ_t=θ_t−1−α·{circumflex over (m)}_t/(√{square root over ({circumflex over (v)}_t)}+ϵ)

where Loss(y_o,c, p_o,c) is a loss function between a network predicted value and a true value, θ is a parameter to be updated in a model, g_t is a gradient obtained by conducting derivation on θ of the loss function f(θ), β₁ is a first-order moment attenuation coefficient, β₂ is a second-order moment attenuation coefficient, m_t is an expectation of the gradient g_t, v_t is an expectation of g_t², {circumflex over (m)}_t is an offset correction of m_t, {circumflex over (v)}_t is an offset correction of v_t, θ_t□1 is a parameter before network update, and θ_t is a parameter after network update.

An actually-shot concrete image is used as a training set and input a trained network, and finally, an classification result of a concrete crack image is output. FIG. 6 shows classification results of some cracks obtained through a tensorboard visualization tool.

With reference to FIG. 1, in the example, the IR7-EC network disclosed herein and current general neural networks including alexnet, googlenet, vgg16, resnet34 and Mobilenet_v3_large that also contains a large number of inverted residual structures are trained separately, and cracks in a concrete crack image test set are identified through the above networks after 20 cycles of training. Precision and loss curves of a training set and a validation set of each model are shown in FIGS. 7A-7D. Final results of a training set, a validation set and a test set of each model are shown in Table 3, where precision and errors of the training set represent a theoretical classification ability of a network after concrete crack training, and precision and errors of the validation set represent an actual classification ability. The higher the precision of the validation set, the lower the error, and the stronger the actual classification ability of the network. A computation formula of precision is as follows:

$\mathrm{accuracy}=\frac{{\sum }_N\mathrm{eq}\left(y_{o,c},\max \left(p_{o,c}\right)\right)}{N}$

• • where y_o,c is a true value of a single picture in a data set (training set/validation set), p_o,c is a network predicted value, which includes 7 probabilities corresponding to 7 crack classes, max( ) is a class corresponding to a value with a greatest probability in extracting p_o,c, eq is used to validate whether a true value (label) y_o,c is equal to max(p_o,c), Σ_N( ) is used to compute the number of images with true values (labels) y_o,c equal to max(p_o,c) in all images in the data set, and N is the number of all crack images in the data set.

A computation formula of loss is as follows:

$\mathrm{Loss}\left(y_{o,c},p_{o,c}\right)=-\sum _{c=1}^M{y}_{o,c}\log \left(p_{o,c}\right)\mathrm{loss}=\frac{{\sum }_{\mathrm{steps}}\mathrm{Loss}\left(y_{o,c},p_{o,c}\right)}{N_{\mathrm{steps}}}{N}_{\mathrm{steps}}=\frac{N}{N_{\mathrm{batch}}}$

• • where Loss(y_o,c, p_o,c) is an error between a network predicted value and a true value of a single image, which is computed through cross entropy, M is the number of classes and is 7 in the patent, N_steps is a step size of network training, N is the number of all crack images in the data set, and N_batch is the number of images included in one batch size and is 16 in the patent.

Table 4 shows the number of parameters and operations of an IR7-EC network and other convolutional neural networks (CNNs), where Flops is the number of floating points of operations, which include all multiplication and addition operations in the network model and used to measure computational complexity of the model. According to results in Table 3, Table 4 and FIGS. 7A-7D, it may be seen that, compared with other networks, concrete crack classification precision and errors of the IR7-EC network disclosed herein are very close to those of Mobilenetv3_large and are greater than those of other compared network models. Meanwhile, for the disclosed network, the number of parameters and operations, a model size and model training time are greatly reduced, and occupied memory is much lower than that of other compared network models except Alexnet. Due to a serious vanishing gradient problem, precision and loss of VGG16 stop changing. With a Resnet34 network as an example, a size of the IR7-EC network model disclosed herein is only 4.09 MB, which is 19.88 times smaller than that of the Resnet34 network. For the disclosed network model, the number of parameters is reduced by 20 times, training time of 20 cycles is shortened by 5.9 times, Flops is reduced by 77.9 times, and meanwhile, higher concrete classification precision and less errors are achieved. With Mobilenetv3_large including inverted residual structures as an example, for the IR7-EC network model disclosed herein, a size is reduced by 3.95 times, the number of parameters is reduced by 4.02 times, training time of 20 cycles is shortened by 1.9 times, Flops is reduced by 4.8 times, and meanwhile, similar concrete classification precision and less errors are achieved.

Reference is made to FIGS. 8A-8F and Table 5 for comparison of classification results of a test set (including 100 crack images of each class). FIGS. 8A-8F show confusion matrices of classification results of test sets of an IR7-EC model and different CNN models, where an ordinate of a matrix being a predicted value, an abscissa of a matrix is a true value, the sum of each row of values represents the number of network predicted samples of the class, and the sum of each column of values represents the number of samples of a true class, such that the larger a value on a diagonal of a matrix, the closer a network predicted value of the class is to a true value. Table 5 shows precision, recall rates and specificity of test sets of an IR7-EC network and other CNN classification networks. The precision is a ratio of samples correctly determined to all predicted positive samples. The higher the precision, the lower a possibility of network reporting errors. A computation formula of precision is as follows:

$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$

The recall (true positive rate) is a ratio of all positive samples predicted correctly to all actual positive samples. The higher the recall rate, the lower a possibility of network reporting missing. A computation formula of the recall rate is as follows:

$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$

The specificity (true negative rate) is a ratio of all negative samples predicted correctly to all actual negative samples. A computation formula of the specificity is as follows.

$\mathrm{Specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$

Reference is made to Table 6 for TP, TN, FP and FN. The second letter includes positive (P) and negative (N), which are used to represent a predicted condition, and the first letter includes true (T) and false (F), which are used to determine an actual condition. Specifically,

• • TP indicates that a network determines that the sample is a positive sample, and the determination is correct (in fact, the sample is positive);
  • TN indicates that a network determines that the sample is a negative sample, and the determination is correct (in fact, the sample is negative);
  • FP indicates that a network determines that the sample is a positive sample, and the determination is false (in fact, the sample is negative); and
  • FN indicates that a network determines that the sample is a negative sample, and the determination is false (in fact, the sample is positive).

With reference to Table 5 and FIGS. 8A-8F, in the test set classification results of all the network models, all the CNNs have a slightly lower ability to classify irregular cracks. In Table 5, except for IR7-EC and Mobilenetv3_large, a recall rate of irregular cracks of the other CNNs is smaller than 0.5. In conclusion, the IR7-EC network model disclosed herein not only has a very light weight, but is comparable to Mobilenetv3_large and is stronger than the other compared CNNs, in test set performance.

TABLE 3
Classification precision and loss values of an IR7-EC
network and other CNNs after 20 iterations
	Precision	Loss value	Test set
	Training	Validation	Training	Validation	precision
Item	set (%)	set (%)	set	set	(%)
IR7-EC	80.5	92.9	0.57	0.23	89.5
Alexnet	80	93.7	0.63	0.25	83.1
VGG16	21.9	27.4	1.91	1.9	14.2
Googlenet	81.3	93	0.95	0.27	73.1
Resnet34	80.9	89.2	0.61	0.32	85.4
Mobilenetv3_large	83.2	95.8	0.52	0.16	90

TABLE 4
The number of parameters and operations of an IR7-EC network and other CNNs
					sum
		Total	Total	weight	time of
	(Total	memory	Flops	file size	training
Item	params)	(MB)	(Flops)	(MB)	(s)
IR7-EC	1,048,448	13.66	47.3M	4.09	765.45
Alexnet	14,595,783	2.77	310.07M	55.6	1526.51
VGG16	134,289,223	109.29	15.5G	512	13230.2
Googlenet	5,980,727	30.03	1.59G	39.4	1689.68
Resnet34	21,288,263	37.61	3.67G	81.3	4521.46
Mobilenetv3_large	4,210,999	50.39	226.44M	16.2	1458.53

TABLE 5
Comparison of precision, recall rates and specificity of test
sets of an IR7-EC network and other CNN classification networks
IR7-EC	Precision	Recall	Specificity	vgg16	Precision	Recall	Specificity
Background	1	1	1	Background	0	0	1
Hole	0.957	0.88	0.993	Hole	0	0	1
IrregularCrack	0.902	0.55	0.99	IrregularCrack	0	0	1
MeshCrack	0.927	0.89	0.988	MeshCrack	0	0	1
ObliqueCrack	0.758	1	0.947	ObliqueCrack	0.143	1	0
TransverseCrack	0.861	0.99	0.973	TransverseCrack	0	0	1
VerticalCrack	0.875	0.91	0.978	VerticalCrack	0	0	1
Mobilenety3large	Precision	Recall	Specificity	googleNet	Precision	Recall	Specificity
Background	1	1	1	Background	1	0.92	1
Hole	0.947	0.9	0.992	Hole	0.421	0.82	0.812
IrregularCrack	0.915	0.65	0.99	IrregularCrack	0.893	0.49	0.99
MeshCrack	0.911	0.92	0.985	MeshCrack	0.907	0.88	0.985
ObliqueCrack	0.824	0.98	0.965	ObliqueCrack	0.719	0.87	0.943
TransverseCrack	0.882	0.97	0.978	TransverseCrack	0.822	0.6	0.978
VerticalCrack	0.885	0.92	0.98	VerticalCrack	0.879	0.58	0.987
resNet34	Precision	Recall	Specificity	AlexNet	Precision	Recall	Specificity
Background	0.99	1	0.998	Background	1	0.95	1
Hole	0.92	0.81	0.988	Hole	0.729	0.78	0.952
IrregularCrack	0.956	0.43	0.997	IrregularCrack	0.885	0.46	0.99
MeshCrack	0.876	0.92	0.978	MeshCrack	0.805	0.95	0.962
ObliqueCrack	0.737	0.98	0.942	ObliqueCrack	0.773	0.99	0.952
TransverseCrack	0.776	0.97	0.953	TransverseCrack	0.873	0.89	0.978
VerticalCrack	0.874	0.9	0.978	VerticalCrack	0.857	0.84	0.977

TABLE Meanings of TP, TN, FP, and FN
	Prediction result
Evaluation index	Positive sample	Negative sample
Actual condition	Positive sample	TP	FN
	Negative sample	FP	TN

The above descriptions are merely preferred examples of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention should fall within the protection scope of the present invention.

Claims

1. A quick and intelligent inverted residual (IR) 7-ECA and CBAM (EC) network based classification method for a concrete image crack type, comprising: building an IR7-EC network model, wherein the IR7-EC network model comprises a convolutional layer, 7 inverted residual-efficient channel attention (ECA) structures, a convolutional block attention module (CBAM) attention mechanism, a pooling layer and a plurality of fully connected layers, wherein the convolutional layer, the 7 inverted residual-ECA structures, the CBAM attention mechanism, the pooling layer and the plurality of fully connected layers are connected sequentially, andeach of the 7 inverted residual-ECA structures is formed by connecting an inverted residual structure and an ECA attention mechanism;inputting an actually-shot concrete image to be detected into a trained IR7-EC network model, and obtaining an initial feature through the convolutional layer;allowing the initial feature to sequentially pass the 7 inverted residual-ECA structures, so as to obtain an enhanced concrete crack feature extraction map;inputting the enhanced concrete crack feature extraction map into the CBAM attention mechanism, and conducting feature extraction at a channel and space level, so as to obtain a feature extraction map containing more crack information; andconducting data dimension reduction and two-dimensional to one-dimensional transformation on the feature extraction map, and classifying an input concrete crack image according to extracted one-dimensional crack image feature information, so as to output an classification result.

2. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 1, wherein the step of building the IR7-EC network model comprises: establishing 13 layers of network frameworks connected sequentially, wherein a 1st layer comprises a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer, and a hardswish activation function, a step size being 2, a number of input channels being 3, and a number of output channels being 16;a 2nd layer to an 8th layer are the 7 inverted residual-ECA structures, and each layer of the 2nd layer to the 8th layer comprises four parts:a first part comprising: a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and an ReLU6 activation function;a second part comprising: a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer, and an ReLU6 activation function;a third part being: an ECA attention mechanism, wherein a structure of the ECA attention mechanism comprises an average pooling layer, a convolutional layer with a convolutional kernel of 1×1, and a sigmoid function; anda fourth part comprising: a convolutional layer with a convolutional kernel of 1×1, a batch normalization layer, and a linear activation function;a 9th layer comprises a convolutional layer with a convolutional kernel of 3×3, a batch normalization layer and a hardswish activation function, a step size being 1, a number of input channels being 96, and a number of output channels being 576;a 10th layer is the CBAM attention mechanism and comprises two parts: a first part being: a channel attention mechanism, wherein the channel attention mechanism comprises an average pooling layer, a maximum pooling layer, a first fully connected layer, an ReLU6 activation function, a second fully connected layer, and a sigmoid function; anda second part being: a space attention mechanism, wherein the space attention mechanism comprises an average pooling layer, a maximum pooling layer, a convolutional layer with a convolutional kernel of 7×7, and a sigmoid function;a 11th layer is an average pooling layer;a 12th layer is a one-dimensional convolutional layer, a hardswish activation function, and a dropout layer with a sparsity ratio of 0.2; anda 13th is a one-dimensional convolutional layer.

3. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 2, wherein training of the IR7-EC network model comprises:collecting crack images, establishing an image data set of different crack types, and conducting pre-processing to obtain a pre-processed image data set;inputting the pre-processed image data set into the 1st layer, and conducting preliminary feature extraction, data normalization and nonlinear transformation on image sample data sequentially through the convolutional layer, the batch normalization layer and the hardswish activation function;inputting a feature preliminarily extracted through the 1st layer into the 2nd layer to the 8th layer, allowing the feature to sequentially pass the 7 inverted residual-ECA structures, inputting the feature into the first part of the inverted residual-ECA structure for processing so as to augment a dimension of an input data channel, inputting data after channel dimension augmentation into the second part so as to conduct deep feature extraction, inputting data after deep feature extraction into the third part, highlighting a plurality of channels with higher network precision by weighting network channels through the ECA attention mechanism, so as to obtain the enhanced concrete crack feature extraction map, inputting the enhanced concrete crack feature extraction map into the fourth part so as to reduce a dimension of a data channel, and allowing the data to pass all the 7 inverted residual-ECA structures sequentially according to similar operations;inputting the data obtained through the 2nd layer to the 8th layer into the 9th layer for feature extraction;inputting data after feature extraction through the 9th layer into the CBAM attention mechanism of the 10th layer, and further conducting feature extraction at the channel and space level on the data, so as to obtain the feature extraction map containing more crack information;inputting the feature extraction map obtained through the 10th layer into the 11th layer, the 12th layer and the 13th layer sequentially so as to conduct data dimension reduction and two-dimensional to one-dimensional transformation, and classifying the input concrete crack image according to the extracted one-dimensional crack image feature information, so as to output the classification result; andcomparing the classification result obtained through the 12th layer and a true result, computing an error through a loss function, and updating a network parameter through an optimization function until a network is fitted.

4. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 2, wherein normalization of the batch normalization layer comprises:

5. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 3, wherein the step of highlighting the plurality of channels with higher network precision by weighting the network channels through the ECA attention mechanism, so as to obtain the enhanced concrete crack feature extraction map comprises:conducting cross-channel interaction on data passing the ECA attention mechanism in each layer, so as to obtain the enhanced concrete crack feature extraction map, by using the following formula:

6. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 3, wherein the step of inputting data after feature extraction through the 9th layer into the CBAM attention mechanism of the 10th layer, and further conducting feature extraction at the channel and space level on the data, so as to obtain the feature extraction map containing more crack information comprises:using average pooling and maximum pooling to aggregate space information of feature mapping, compressing a space dimension of an input feature map, and conducting summation and merging element by element, so as to generate a channel attention map, through the following formula: Mc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))wherein Mc represents channel attention, MLP( ) is composed of the first fully connected layer, the ReLU6 activation function, and the second fully connected layer, σ is a sigmoid operation, F is an input feature map, AvgPool( ) is average pooling, MaxPool( ) is maximum pooling, and Ms represents the space attention mechanism; andusing average pooling and maximum pooling to compress the input feature map in a space attention module, so as to obtain the feature extraction map containing more crack information through the following formula: MS(F)=σ(f7*7[AvgPool(F),MaxPool(F)])wherein MS represents the space attention mechanism, σ is a sigmoid operation, f7*7[⋅] represents a convolutional operation of 7*7, F is the input feature map, AvgPool( ) is average pooling, and MaxPool( ) is maximum pooling.

7. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 3, further comprising: conducting sparsification on data passing the dropout layer in each layer, so as to avoid network over-fitting through the following formulas: rj(l)˜Bernoulli(p){tilde over (y)}(l)=r(l)*y(l) wherein a Bernoulli(p) function is used to generate a vector of a probability rj(l), such that a neuron stops working with probability p, y(l) is an upper layer of an output feature map, and {tilde over (y)}(l) is a feature map output after passing the dropout layer.

8. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 3, wherein the step of computing the error through the loss function comprises:computing loss of the network through the following formula:

9. The quick and intelligent IR7-EC network based classification method for the concrete image crack type according to claim 3, further comprising: optimizing internal parameters of the network through the following formulas: f(θ)=Loss(yo,c,po,c)gt=∇θft(θt−1)mt=β1·mt−1+(1−β1)·gt vt=β2·vt−1+(1−β2)·gt2 {circumflex over (m)}t=mt/(1−β1t){circumflex over (v)}t=vt/(1−β2t)θt=θt−1−α·{circumflex over (m)}t/(√{square root over ({circumflex over (v)}t)}+ϵ)wherein Loss(yo,c, po,c) is a loss function between a network predicted value and a true value, θ is a parameter to be updated in a model, gt is a gradient obtained by conducting derivation on θ of the loss function f(θ), β1 is a first-order moment attenuation coefficient, β2 is a second-order moment attenuation coefficient, mt is an expectation of the gradient gt, vt is an expectation of gt2, {circumflex over (m)}t is an offset correction of mt, {circumflex over (v)}t is an offset correction of vt, θt−1 is a parameter before network update, and θt is a parameter after network update.