Assessing compressive strengths of mortar and concrete from digital images by machine learning techniques

Abstract

Compressive strength is the most important metric of concrete quality. Various nondestructive and semi-destructive tests can be used to evaluate the compressive strength of concrete. In the present study, a new image-based machine learning method is used to predict concrete compressive strength, including evaluation of six different models. These include support-vector machine model and various deep convolutional neural network models, namely AlexNet, GoogleNet, VGG19, ResNet, and Inception-ResNet-V2. In the present investigation, cement mortar samples were prepared using each of the cement:sand ratios of 1:3, 1:4, and 1:5, and using the water:cement ratios of 0.35 and 0.55. Cement concrete was prepared using the cement:sand:coarse aggregate ratios of 1:5:10, 1:3:6, 1:2:4, 1:1.5:3 and 1:1:2, using the water:cement ratio of 0.5 for all samples. The samples were cut, and several images of the cut surfaces were captured at various zoom levels using a digital microscope. All samples were then tested destructively for compressive strength. The images and corresponding compressive strength were then used to train machine learning models to allow them to predict compressive strength based upon the image data. The Inception-ResNet-V2 models exhibited the best predictions of compressive strength among the models tested. Overall, the present findings validated the use of machine learning models as an efficient means of estimating cement mortar and concrete compressive strengths based on digital microscopic images, as an alternative nondestructive/semi-destructive test method that could be applied at relatively less expense.

Graphical abstract

Keywords

support vector machine / deep convolutional neural network / microscope / digital image / curing period

1 Introduction

Concrete is near to the most widely used material worldwide, second only to water. The raw ingredients of cement concrete are cement, fine aggregate, coarse aggregate, and water, whereas mortar is prepared by mixing cement with fine aggregate and water only.

Strength, workability, and durability are the most important properties of concrete, and depend greatly on the amounts of raw ingredients used and their quality. Compressive strength is the primary factor tested to estimate the strengths of cement mortar and concrete. Compressive strength is the most important parameter for designing reinforced concrete structures and is also a key indicator used to monitor the health of existing structures. Therefore, evaluating the compressive strengths of mortar and concrete is vital for both existing and new reinforced concrete (RC) structures. This is made challenging by the fact that concrete is a heterogeneous material and thus it is very difficult to accurately predict any of its properties.

Various tests are used to determine compressive strength for cement mortar or concrete. Cube testing is conducted in the laboratory on fresh concrete samples made using the same amount of raw ingredients used in new concrete members. It is very helpful for predicting the strength of newly made concrete members but is very difficult to apply to predict the strength of existing structures. To estimate the strength of various components of existing structures, various other tests are conducted. Ultrasonic pulse velocity, rebound hammer, and sonic rebound (SonReb) tests are examples of tests that can be applied nondestructively on site [1]. However, these tests are not highly accurate for predicting the actual compressive strength of concrete. Core cutter tests yield accurate results but are destructive in nature which sometimes heavily damages the concrete member [2]. Overall, all these methods require costly and sophisticated equipment with high levels of maintenance, and require proficient and specialized workforce with high degrees of expertise and integrity.

Artificial neural network (ANN) is a type of machine learning (ML) algorithm inspired by the neurons of the human brain and is a widely used method for data prediction [1–4]. Unlike the human brain, a neural network also consists of neurons, also called nodes, and by several means, the nodes are inter-connected. A neural network combines several processing layers, using simple elements operating in parallel and inspired by the biological nervous system. It is to be noted that ANN consists of three layers—input layer, hidden layer, and output layer. The input layer receives the inputs, hidden layer processes the inputs, and output layer generates the result. Actually, each layer tries to acquire certain weights. In a Feed-Forward Neural network, inputs are processed only in the forward direction. In ANN, activation functions introduce nonlinear properties to the network which helps the network to learn any complex relationship between input and output. It is to be noted that ANN can be used to solve problems related to tabular data, image data, and text data. However, while solving an image classification problem using ANN, the first step is to convert a two-dimensional image into a one-dimensional vector prior to training the model. For this purpose, the number of trainable parameters increases considerably with an increase in the size of the image. In addition, ANN misses the spatial structures of an image like the orientation of pixels of the image. To avoid these types of difficulties, convolutional neural networks (CNN) models are being used for various applications and domains, and are particularly predominant in image and video processing projects [5]. In CNN, Kernels are used to extract the needed information from the input using the convolution operation. CNN learns the filters automatically without mentioning it explicitly. The filters support in extracting the right and relevant information from the input data. Further, CNN captures the spatial information from an image, i.e., the arrangement of pixels and the connection between them in an image. Thus, CNN is extremely useful for reducing the hyper parameters and increasing the image information extraction.

The use of ML techniques for image classification is promising for this type of problem [6,7]. Commonly used ML techniques include support-vector machines (SVMs), ANN and deep convolutional neural networks (DCNNs). ML has been applied to various problems such as recognition of handwritten digits [8], face recognition [9], pest detection [10], recognition of plant disease [11,12], autonomous driving [13], and medical diagnosis [14–16]. In civil engineering, various ML methods have been used on image data to identify pavement cracks [17,18], recognize structural damage [19], identify concrete cracks [20], etc. Recently these image processing technique(s) were applied by some esteemed researchers for predicting the compressive strength of concrete. It is important to mention that this method is nondestructive in nature and requires less time and cost. Researchers have carried out statistical analysis and ANN analysis of captured images and image processing data of concrete samples, and compared the results to destructive experimental test results of compressive strength. On the basis of concrete image data, Başyiğit et al. [21] carried out several types of regression analyses to predict the compressive strength of concrete. Dogan et al. [22] used an ANN to evaluate concrete compressive strength based upon concrete image processing data. However, among ML techniques, DCNN is the most modern and effective tool for image classification. The efficacy of DCNN methods depends on their architecture, including the number, type, and size of layers used. In recent years several DCNN architectures have been developed for image classification, including AlexNet, GoogleNet, VGG19, ResNet, and Inception-Resnet-V2. Jang et al. [23] successfully used these modern DCNN architectures for predicting concrete compressive strength on a few samples of cement concrete taken using a digital microscope at a particular resolution, and reported that ResNet yielded satisfactory results. However, this work did not extend to the use of DCNNs for both cement mortar and concrete prepared by using different raw materials. In addition, this work did not include the use of Inception-Resnet-V2, which has proved the most effective DCNN for classifying images.

Accordingly, in the present work we extended the use of DCNN image analysis to predict the compressive strengths of cement mortars and concretes of various compositions, including the use of Inception-ResNet-V2. Samples of cement mortar and of concrete materials were prepared, using various proportions of cement, sand, and water for cement, and of cement, sand, coarse aggregate, and water for concrete. The samples were cured for various durations and images of the samples were captured at various zoom levels using a digital microscope, and the samples were tested destructively for compressive strength. Then, a SVM model and five different DCNN models, namely AlexNet, GoogleNet, VGG19, ResNet, and Inception-ResNet-V2, were applied to predict the compressive strengths of the materials based upon their images. The accuracy of each ML method was evaluated in terms of R² and root mean square error (RMSE) values of the relations modeled between predicted and measured compressive strength.

2 Materials and methods

This section describes the details of the SVM and DCNNs used to predict the concrete compressive strength based on the concrete image data. The present investigation was focused on finding the most suitable DCNN model. The experimental process was divided into four basic steps: determining the compressive strength of cement mortar and concrete, preprocessing of image, training of image, classification of image, and evaluation of the result. The detailed procedure are presented below.

2.1 Experimental program

Mortar cubes were prepared using the cement: sand ratios of 1:3, 1:4 and 1:6, and further using two different water:cement ratios for each: 0.35 and 0.55. Concrete cubes were prepared using the cement: sand: coarse aggregate ratios of 1:5:10, 1:3:6, 1:2:4, 1:1.5:3, and 1:1:2, using the water:cement ratio of 0.5 for all samples. For each mix proportion of both mortar and concrete, eighteen cylindrical samples having diameter of 100 mm and length of 200 mm have been prepared. Among each of these, three samples have been tested for compressive strength in compression testing machine (CTM) at 7, 28, and 56 d of curing and their average compressive strength has been taken for each curing period. The tests have been conducted according to relevant IS codes. Additionally, three samples both of mortar and concrete were cut into pieces after scheduled curing and their images were captured. Fig.1 illustrates some of the samples, their preparation, and their strength testing.

2.2 Image acquisition and image data preprocessing

Images of mortar and concrete were captured. Fig.2 shows mortar and concrete samples prepared for image capture, and the digital microscope used to capture the images. For mortar, the arbitrary number of ten images were captured for each sample condition; images were captured at 1×, 20×, and 40× zoom. Dewinter Caliper Pro software was used to capture, process, and store images. Thus, for mortar, 60 total images were captured at each zoom level and at each stage of curing. Similarly, 50 images were captured of concrete at each zoom level and at each stage of curing. In the present study both random cropping and horizontal flipping were conducted for all the images. Particularly, an image with an original size of 1600 × 1200 was scaled to 112 × 112. Further, a random seed was created in the 18 × 18 portion. Furthermore, 84 × 84 segment of the image was chosen using the random seed. The resulting 84 × 84 image may also be rotated 180° about the horizontal axis. Finally, the 9 sets of mortar images and 9 sets of concrete images, each containing 120 or 100 images, were divided into a training set of 90%, and a testing set of 10%. Fig.3 shows the images captured of mortar having the cement:sand ratio of 1:3 and water:cement ratio of 0.35, and of concrete having the cement:sand:aggregate ratio of 1:1.5:3, at each of the three levels of zoom and three curing periods.

2.3 Model training

Training of the learning models was conducted using the dataset described above. Several ML tools are available for solving image classification problems, among which SVM and DCNN are the most commonly used. The background theory of these techniques are mentioned below.

2.3.1 Support vector machines

SVMs can be used for tasks including estimation, regression, classification, and structural risk reduction [5,7] [8,9]. It yields a data set \{\left(y_1,{\mathit{\text{z}}}_1\right),...,\left(y_n,{\mathit{\text{z}}}_n\right)\}, where y represents an input variable, z represents a target output and n represents number of training samples. Here, y is called a support vector. The aim of SVM is to create an analytical model to establish the relationship between the input and output variables by means of a nonlinear mapping procedure. Here, nonlinear mapping function is created with a connection weight vector and the bias, determined by solving an optimization problem using objective function. It is to be noted that by introducing the kernel function, the nonlinearity of the SVM can be resolved easily [6].

2.3.2 Deep convolutional neural network

DCNN has proven to be very advantageous in image classification performance in recent years. DCNN is a deep learning algorithm that is intended to develop the image data obtained from multiple arrays by extracting particular features instead of data distortions arising from shifting, rescaling, and noise [24]. A DCNN mainly comprises convolutional layers, pooling layers and fully connected layers [25] (Fig.4). It is to be noted that the convolutional layer, tailed by a pooling layer comprises of a number of convolution kernels which are introduced to determine various feature maps. It is also to be noted that the element involved in carrying out the convolution operation in the Convolutional Layer is called the Kernel/Filter. The mth feature of the nth layer of the convolutional, y_m^n, can be formulated as follows:

(1)a_m^n= \sum _{j=1}^{N^{n-1}}{K}_{jm}^n\ast y_j^{n-1}+ B_m^n, y_m^n= {f(a}_m^n).

In this equation, y_j^{n-1} represents the jth feature map of (n-1 )th layer, N^{n-1} represents the number of feature maps of the layer, K_{jm}^n represents the convolutional kernel related to the mth map between the nth layer and the jth map in the (n-1)th layer, and B_m^n represents the bias of the kernel as mentioned in Eq. (1), f(\cdot ) [25]. This explains the element-wise nonlinear activation function which is applied on nonlinearity into the multilayer systems. There are several activation functions such as sigmoid, tanh, and rectified linear unit (ReLU) [26]. After conducting the convolution operation in the convolution layer, the pooling layer integrates features, decreases parameters, and changes the invariance by lessening the feature map resolution. Representing the pooling operation as downsample(\cdot ), for each feature map y_m^n, it can be written as

(2)s_m^n= \mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}\mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\left(y_m^n\right).

Typical pooling operations applied are average pooling and max pooling. The pooling operation considers some k× k area and yields a single value by computing the mean or maximum in that area.

After the abovementioned convolutional and pooling layers, in some cases, several fully connected layers are applied (e.g., AlexNet, LeNet, VGG). The objective of a connected layer is to learn the feature maps at the mid-level. Note that a huge number of weight parameters are needed to implement a full connection. As in ANN, a feed-forward or feed-backward process can be applied. However, most recently developed networks use a forward process, which can be expressed as

(3)a_m^n= \sum _{j=1}^{N^{n-1}}{y}_j^{n-1}{W}_{jm}^n+B_m^n, y_m^n= {f(a}_m^n),

where B_m^n and W_{jm}^n respectively represent the bias term and weight vector at the jth filter of the mth layer. The final layer of a neural network is usually a Softmax layer, which transforms the output of the end layer to the necessary probability distribution. It is used for estimating the input data class label. Consider A_j and {\mathit{\text{z}}}_j, denoting the estimated label and actual label for the input sample. The loss function is:

(4)E=\frac{1}{2}\sum _{j=2}^{N^L}\parallel {{\mathit{\text{z}}}_j-A_j\parallel }^2,A_j=y_j^L,

where N^L denotes the number of nodes of the Lth layer and E denotes the total error in classification of all the output points for a particular sample. The loss function of Eq. (4) is termed Euclidean loss, though other losses can be applied such as sigmoid loss, contrastive loss, Softmax loss, information gain loss, hinge loss, and entropy loss [27].

Transfer learning is a type of research topic in ML that aims on storing information obtained while resolving one problem and applying it to a dissimilar but associated problem [3,11–13]. Transfer learning can be defined in terms of domains and tasks. A domain \delta comprises of a feature space \mathit{y} and a marginal probability distribution \mathit{P}\left(Y\right), where \mathit{Y}=\left\{{\mathit{y}}_{\mathbf{1}},...,{\mathit{y}}_{\mathit{n}}\right\}\in \mathit{y}. A task comprises of a label space \mathit{\text{z}} and an objective predictive function \mathit{f}: \mathit{y}\to \mathit{\text{z}} for a particular domain \delta =\{\mathit{y},\mathit{P}\left(Y\right)\}. The function f is used to assess the related label \mathit{f}\left(\mathit{y}\right) of new instance y. The task (\mathit{\tau }=\left\{\mathit{\text{z}},f\left(\mathit{y}\right)\right\}) is trained from the training data comprising of sets {{\mathit{y}}_{\mathit{i}}, {\mathit{\text{z}}}_{\mathit{i}}\}, where {\mathit{y}}_i\in \mathit{Y} and {\mathit{\text{z}}}_i\in \mathit{\text{z}} . It is to be noted that for a source domain {\delta }_{\mathrm{s}} and training task {\mathit{\tau }}_{\mathrm{s}}, a target domain {\delta }_{\mathrm{T}} and training task {\mathit{\tau }}_{\mathrm{T}}, ({\delta }_{\mathrm{s}}\ne {\delta }_{\mathrm{T}}, or {\mathit{\tau }}_{\mathrm{s}}\ne {\mathit{\tau }}_{\mathrm{T}}) transfer training focuses to increase the training of the target projecting function f_T(·) in {\delta }_{\mathrm{T}} utilizing the information in {\delta }_{\mathrm{s}} and {\mathit{\tau }}_{\mathrm{s}} .

Several researchers have proposed several DCNN models in recent years; these are discussed below.

1) AlexNet

AlexNet, developed by Krizhevsky et al. [28], is a CNN model for large-scale image classification. AlexNet has 60 million parameters, and comprises 5 convolutional layers, 3 pooling layers, and 3 fully connected layers. The model uses 227 × 227 × 3 image inputs. The first layer is a convolution network comprising 11 × 11 filters, 96 of which are applied at stride 4 with output volume 55 × 55 × 96. The second layer is a pooling layer comprising 3 filters of size 3 × 3 which are applied at stride 2. The output of the pooling layer is 27 × 27 × 96, having no parameters to learn. In the end, the network output meets a chain of two numbers of fully connected layers of size 4096. The last layer is a Softmax layer comprising 1000 ImageNet classes. The second fully connected layer is a Softmax classifier. Further, it is to be noted that, to avoid overfitting in fully connected layers, the designers applied a regularization technique called dropout, considering a ratio of 0.5. Additionally, the ReLU function is applied to each of the first seven layers in the AlexNet model. The network is illustrated schematically in Fig.5.

2) VGG19

VGG19 [29] is a 19-layer DCNN, comprising six main structures each of which generally comprises multiple connected convolutional layers and fully connected layers. The convolutional kernel is 3 × 3, and the input is 224 × 224 × 3. The pretrained network can classify pictures into 1000 object categories. The details of the network are presented in Fig.6. Though it is an extremely simple structure having a simple linear chain of layers, with all the convolutional layers included it is capable of making highly accurate classifications.

3) GoogleNet

GoogleNet was published by Szegedy et al. [30-32] in 2017; it comprises seven million parameters and encompasses 9 inception modules (Fig.7), 4 convolutional layers, 4 max pooling layers, 3 average pooling layers, 5 fully connected layers, and 3 Softmax layers. Dropout regularization is applied in the fully connected layer and the ReLU activation function is used in each convolutional layer. Though this network is wider and deeper than AlexNet, having 22 layers total, it uses fewer network parameters (7 million).

4) Residual Network (ResNet)

The ResNet model [31] is based on deep architectures and shows good convergence with satisfactory accuracy. The main base element of ResNet is the residual block, comprising numerous weighted residual units having different numbers of layers. These layers operate on top of each other, and each layer is intended to acquire knowledge from some essential mapping having an intended function. However, the quantity of the operations may be different based on variations of architectures. The residual units each consist of a convolutional layer and pooling layers. ResNet has a very deep architecture, with 25.6 million parameters and enabling up to 50 layers stack depth all the layers periodically. The model ends with fully connected layers having 1000 output classes. ResNet is illustrated schematically in Fig.8.

5) Inception-ResNet-V2

Inception-ResNet-V2 [32,33] is a DCNN having 164 layers depth and 55 million parameters. It can categorize images into 1000 object categories. For this reason, the network has learned rich feature representations for a wide range of images. The image input dimensions are 299 × 299 and the output is a list of predicted class probabilities. Fig.9 schematically illustrates the Inception-ResNet-V2 model.

In the present work, Matlab 2021 software was used to implement all the ML methods tested. This software is capable of the task, but requires a very high computational cost and a high-performance workstation. The Intel Xeon E5-2620 2.1 GHz; RAM: 64 GB was used.

3 Results and discussion

To establish a relationship between concrete image data and compressive strengths of mortar and concrete, compressive strength was first measured experimentally. For each mix proportion and each curing duration, three samples were prepared and their strengths were averaged. Tab.1 lists the results. As the amount of cement was increased, the compressive strength was observed to increase, whereas the compressive strength decreased significantly with increasing sand and water contents. Unsurprisingly, compressive strength increased with increasing curing time.

The SVM and DCNN models were trained and then applied to relate image data to compressive testing results. Tab.2 shows hyper parameter used in different DCNN models. The graphical presentation of loss function of training and validation dataset for concrete and mortar at 28 d curing, capture picture at 20× zoom, is shown in Fig.10. The results of predicted vs. actual compressive strengths of mortar and concrete for each ML model tested are presented in Fig.11. It is pertinent to mention that twelve number of samples of mortar and concrete at each above-mentioned mix proportion were tested each for 7, 28, and 56 d curing period at original, 2× and 4× zoom for all SVM, AlexNet, GoogleNet, VGG19, ResNet, and Inception-ResNet-V2. Thus, the total sample points are 648 for mortar and 540 for concrete. Fig.11(a) and 11(b) clearly depict that most of the data points are well predicted in that they are situated on center line of the graphs. It is to be noted that the huge number of data points lies one on another at centerline. Few points which are not well predicted are not at centerline as well. Further, to determine model precision, R² and RMSE were computed for each set of results.

The coefficient of R^2 and RMSE can be defined as

(5)R^2= \frac{{\left(n\sum _i^n{y}_i{y}_i^{\prime }-\sum _i^n{y}_i^{\prime }\sum _i^n{y}_i\right)}^2}{\left(n\sum _i^n{y}_i^{\prime 2}-{\left(\sum _i^n{y}_i^{\prime }\right)}^2\right)\left(n\sum _i^n{y}_i^2-{\left(\sum _i^n{y}_i\right)}^2\right)},

(6)RMSE= \sqrt{\frac{1}{n}{\sum }_i^n{\left(y_i-y_i^{\prime }\right)}^2},

where y_i and y_i^{\prime } are the actual and predicted compressive strength respectively. n is the number of data sample. Note that R² signifies how well the independent variable is related to the dependent variable, with R² closer to 1 signifying better prediction power. RMSE represents the absolute difference of actual and predicted compressive strength, i.e., error in prediction of the compressive strength. The R² and RMSE metrics of the relations between predicted and actual compressive strength of cement mortar and concrete for the various ML models studied are charted in Fig.12 and Fig.13. It is observed that SVM yielded poor results in terms of R² (R² < 0.7) as expected for both mortar and concrete. Among the five DCNN models studied, the accuracy varied because of the differences in model architectures. It is to be noted that Inception-ResNet-V2 produced the best strength estimates for both mortar and concrete. Further, the less improved architecture of AlexNet, AlexNet yielded the lowest R² values among all DCNN models used for predicting strength for both mortar and concrete. Inception-ResNet-V2 also produced the smallest RMSEs and SVM produced the largest. It can be concluded that Inception-ResNet-V2 performed the best in predicting mortar and concrete compressive strengths from image data. Earlier, Jang et al. [23] reported that ResNet yielded the best results (R² = 0.764). However, Inception-ResNet-V2, which combines ResNet and GoogleNet, yielded the best results in the present work (with R² values close to 0.9 for mortar and 0.85 for concrete). Overall, the models predicted the strengths of cement mortar more accurately than those of cement concrete. This is likely because mortar is more homogeneous than concrete and thus more predictable. The variations in image zoom and curing period did not appear to affect the estimation results excessively. Furthermore, the RMSEs obtained in the present work were much lower than those obtained by Jang et al. [23]. It is also notable that the use of pictures captured at different zoom levels helps to improve estimation accuracy. Overall, the present study clearly demonstrated that DCNN ML models are capable of extracting multifaceted discriminative features from concrete images. These results validate the potential efficacy and applicability of the present method for predicting concrete compressive strengths of mortar and concrete. Thus, the DCNN image analysis method has the potential to be applied as a nondestructive in situ test with comparatively low cost.

4 Conclusions

Compressive strength is the most important metric of concrete quality. Due to their heterogeneity, it is very difficult to predict the compressive strengths of both cement mortar and concrete. In the present study, we investigated the use of ML techniques to analyze images of these materials to predict their compressive strength. Cement mortar samples were tested having various contents of cement, sand, and water, and likewise concrete samples were tested having various contents of cement, sand, coarse aggregate, and water. Samples were cured for various durations and images were captured at various levels of zoom. A SVM model and five different DCNN models were applied to establish relationships between the captured image parameters and compressive strengths. Ninety percent of the images were used as a training set, and the remaining ten percent were used for testing. Among all methods tested, Inception-ResNet-V2 clearly yielded the best predictions of compressive strength of both materials, giving the highest R² values (of about 0.9) and the lowest RMSEs. In the present investigation, Inception-ResNet-V2 performed extremely well and represented an unconditional improvement upon the result obtained by ResNet.

Due to the limited variety in the range of compressive strengths of the samples, and in the images acquired, the models developed in the current investigation studied are likely to yield erroneous compressive strength results for new images representing samples whose compressive strengths are not within the input range. Further, K fold cross validation technique make the training procedure more reliable. It is obvious that one sampling comparison is not convincing enough. However, it needs extensive time to conduct this cross validation process. Thus, this technique was not applied in the present study. Nevertheless, this method can be successfully applied in the field for predicting the quality of mortar and concrete. It is important to mention that this method is nondestructive in nature and requires less time and cost than existing methods. However, expert workmanship is still required to capture suitable images on site and to analyze them.

References

[1] Breysse D, Romão X, Alwash M, Sbartaï Z M, Luprano V A M. Risk evaluation on concrete strength assessment with NDT technique and conditional coring approach. Journal of Building Engineering, 2020, 32 : 101541

[2] Jafari S, Rots J G, Esposito R. Core testing method to assess nonlinear shear-sliding behaviour of brick-mortar interfaces: A comparative experimental study. Construction & Building Materials, 2020, 244 : 118236

[3] Guo H, Zhuang X, Rabczuk T. A deep collocation method for the bending analysis of Kirchhoff plate. Computers, Materials & Continua, 2019, 59( 2): 433–456

[4] Anitescu C, Atroshchenko E, Alajlan N, Rabczuk T. Artificial neural network methods for the solution of second order boundary value problems. Computers, Materials & Continua, 2019, 59( 1): 345–359

[5] Herzog S, Tetzlaff C, Wörgötter F. Evolving artificial neural networks with feedback. Neural Networks, 2020, 123 : 153–162

[6] Haftkhani A R, Abdoli F, Sepehr A, Mohebby B. Regression and ANN models for predicting MOR and MOE of heat-treated fir wood. Journal of Building Engineering, 2021, 42 : 102788

[7] Martini R, Carvalho J, Arêde A, Varum H. Validation of nondestructive methods for assessing stone masonry using artificial neural networks. Journal of Building Engineering, 2021, 42 : 102469

[8] Niu X X, Suen C Y. A novel hybrid CNN-SVM classifier for recognizing handwritten digits. Pattern Recognition, 2012, 45( 4): 1318–1325

[9] Derman E, Salah A A. Continuous real-time vehicle driver authentication using convolutional neural network based face recognition. In: 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018). Xi’an: IEEE, 2018, 577–584

[10] Ebrahimi M, Khoshtaghaza M, Minaei S, Jamshidi B. Vision-based pest detection based on SVM classification method. Computers and Electronics in Agriculture, 2017, 137 : 52–58

[11] Arnal Barbedo J G. Digital image processing techniques for detecting, quantifying and classifying plant diseases. SpringerPlus, 2013, 2( 1): 1–12

[12] Xu G, Zhang F, Shah S G, Ye Y, Mao H. Use of leaf color images to identify nitrogen and potassium deficient tomatoes. Pattern Recognition Letters, 2011, 32( 11): 1584–1590

[13] Nugraha B T, Su S F. Towards self-driving car using convolutional neural network and road lane detector. In: 2017 2nd International Conference on Automation, Cognitive Science, Optics, Micro Electro-mechanical System, and Information Technology (ICACOMIT). Jakarta: IEEE, 2017, 65–69

[14] Sun W, Tseng T L B, Zhang J, Qian W. Enhancing deep convolutional neural network scheme for breast cancer diagnosis with unlabeled data. Computerized Medical Imaging and Graphics, 2017, 57 : 4–9

[15] Dabeer S, Khan M M, Islam S. Cancer diagnosis in histopathological image: CNN based approach. Informatics in Medicine Unlocked, 2019, 16 : 100231

[16] Parashar J, Sumiti M. Breast cancer images classification by clustering of ROI and mapping of features by CNN with XGBOOST learning. Materials Today: Proceedings, 2020, 1–9

[17] Gopalakrishnan K, Khaitan S K, Choudhary A, Agrawal A. Deep Convolutional Neural Networks with transfer learning for computer vision-based data-driven pavement distress detection. Construction & Building Materials, 2017, 157 : 322–330

[18] Hoang N D, Nguyen Q L. A novel method for asphalt pavement crack classification based on image processing and machine learning. Engineering with Computers, 2019, 35( 2): 487–498

[19] Lin Y Z, Nie Z H, Ma H W. Structural damage detection with automatic feature-extraction through deep learning. Computer-Aided Civil and Infrastructure Engineering, 2017, 32( 12): 1025–1046

[20] Cha Y, Choi W, Büyüköztürk O. Deep learning-based crack damage detection using convolutional neural networks. Computer-Aided Civil and Infrastructure Engineering, 2017, 32( 5): 361–378

[21] Başyiğit C, Çomak B, KilinçarslanŞ, SerkanÜncü I. Assessment of concrete compressive strength by image processing technique. Construction & Building Materials, 2012, 37 : 526–532

[22] Dogan G, Arslan M H, Ceylan M. Concrete compressive strength detection using image processing based new test method. Measurement, 2017, 109 : 137–148

[23] Jang Y, Ahn Y, Kim H Y. Estimating compressive strength of concrete using deep convolutional neural networks with digital microscope images. Journal of Computing in Civil Engineering, 2019, 33( 3): 04019018

[24] Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998, 86( 11): 2278–2324

[25] Han D, Liu Q, Fan W. A new image classification method using CNN transfer learning and web data augmentation. Expert Systems with Applications, 2018, 95 : 43–56

[26] SimonyanKZisserman A. Very deep convolutional networks for large-scale image recognition. 2015, arxiv: 1409.1556

[27] Lowe D G. Object recognition from local scale-invariant features. Proceedings of the seventh IEEE International Conference on Computer Vision, 1999, 2 : 1150–1157

[28] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 2017, 60( 6): 84–90

[29] Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, Berg A C, Fei-Fei L. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 2015, 115( 3): 211–252

[30] Mathworks. GoogLeNet convolutional neural network—MATLAB googlenet—MathWorks Benelux. 2021 (available at the website of Mathworks)

[31] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Las Vegas, NV: IEEE, 2016, 770–778

[32] Szegedy C, Ioffe S, Vanhoucke V, Alemi A A. Inception-V4, Inception-ResNet and the impact of residual connections on learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, CA: AAAI Press, 2017, 4278–4284

[33] Mathworks. Pretrained Inception-ResNet-v2 convolutional neural network—MATLAB inceptionresnetv2—MathWorks América Latina. 2021 (available at the website of Mathworks)

RIGHTS & PERMISSIONS

Higher Education Press