Surrounding rock classification from onsite images with deep transfer learning based on EfficientNet

Abstract

This paper proposes an accurate, efficient and explainable method for the classification of the surrounding rock based on a convolutional neural network (CNN). The state-of-the-art robust CNN model (EfficientNet) is applied to tunnel wall image recognition. Gaussian filtering, data augmentation and other data pre-processing techniques are used to improve the data quality and quantity. Combined with transfer learning, the generality, accuracy and efficiency of the deep learning (DL) model are further improved, and finally we achieve 89.96% accuracy. Compared with other state-of-the-art CNN architectures, such as ResNet and Inception-ResNet-V2 (IRV2), the presented deep transfer learning model is more stable, accurate and efficient. To reveal the rock classification mechanism of the proposed model, Gradient-weight Class Activation Map (Grad-CAM) visualizations are integrated into the model to enable its explainability and accountability. The developed deep transfer learning model has been applied to support the tunneling of the Xingyi City Bypass in the high mountain area of Guizhou, China, with great results.

Graphical abstract

Keywords

surrounding rock classification / convolutional neural network / EfficientNet / Gradient-weight Class Activation Map

1 Introduction

The rock mass is a complex structure formed in various long-lasting and complex geological evolution processes, resulting in various rock mass structural planes, including joints, fissures, bedding, fault fracture zones, etc. Due to the variability, inhomogeneity, and uncertainty of rock mass and the concealment of geological body, the unpredictability of rock destruction is difficult to control. Making a correct evaluation of the surrounding rock grade during tunnel excavation is not only related to the design, construction plan and cost control of the tunnel but also to the safety and stability during the construction, operation, and maintenance, which is undoubtedly a severe challenge for scholars and engineers.

With the fast development of data availability, computing power and abundant open-source platforms, deep learning (DL) has been widely applied in computer vision [1]. Compared with traditional neural networks, deep convolutional networks have hundreds of neurons per layer and up to five convolutional stages, which have improved the specific applications on some standard classification data sets [2–4]. Thus, in tunneling, image recognition of tunnel face information has attracted more and more scientists and engineers [5]. So far, several scholars have used different methods to determine the structural mechanism of different rock masses. For example, Aksoy et al. [6] used the classic field test method to compare different empirical equations for the rock modulus by combining displacement on-site measurement and numerical modeling. Goh and Zhang [7] determined the safety factor for caverns under different dimensions and rock strength using finite difference method. Cha et al. [8] proposed a structural visual inspection method based on the Faster Region-based convolutional neural network (CNN), which realized the quasi-real-time simultaneous detection of multiple types of damage. Ran et al. [9] developed an accurate field rock type identification approach based on deep CNN image analysis that can identify six common rock types with 97.96% overall classification accuracy, outperforming other established DL and linear models. Chen et al. [10] implemented a CNN model to recognize multiple rock structures based on tunnel face images with Inception-ResNet-V2 (IRV2), classifying 5 types of rock structures: mosaics, granules, layers, blocks, and fragments. However, these methods are very laborious. To train a good model, excellent hardware and a lot of training time are required, and the training images they used are carefully prepared, which is not practical in actual engineering. Therefore, we focus on the efficiency and accuracy of the surrounding rock classification from onsite images, developing a model which can be used in practical engineering. This may bring a new direction for the application of DL in tunnel engineering.

Compared with the traditional surrounding rock classification method, DL can reduce the influence of subjective judgment, save the operation costs and improve the classification efficiency. In DL, CNN, as an important class of artificial neural networks in image processing, is widely used in geotechnical engineering. Image acquisition and computer image pattern recognition technology enable a few automated systems to recognize rock images [11–13]. Those systems have greatly assisted engineers to improve accuracy and efficiency of identification. Surrounding rock image classification using DL has become a research hotspot. Patel and Chatterjee [14] developed a vision-based model on a laboratory scale using a probabilistic neural network. The input features are color histograms. In combination with remote sensing, Kowk et al. [15] proposed a CNN model for deriving a high-definition map of naturally exposed rocks throughout Hong Kong. Based on DL, Huang et al. [5] proposed a novel image recognition algorithm for semantic segmentation of metro shield tunnel cracks and leaks using fully convolutional network-extracted feature hierarchies with excellent results. Using TensorFlow, Alférez et al. [16] developed several CNN models for classifying granitoids, which are the most popular plutonic rocks in the earth’s crust. Unlike conventional manual analysis, Cheng and Guo [17] proposed a CNN for thin section granularity analysis by automatically extracting features from image samples and building image granularity recognition classifier. To identify lithological patterns in Brazilian pre-salt carbonate rocks, dos Anjos et al. [18] developed an automated workflow to accelerate the lithology classification from microtomographic images based on four CNN models. To improve the classification performance on a small sample number of fine-grained rocks, Liang et al. [19] proposed a fine-grained image classification network combining image cutting method and SBV algorithm. Li et al. [20] achieved the classification accuracy of 100% on 4 groups of Martian rock images collected from MSL Analyst’s Notebook, with a VGG-16 based architecture [21] with deep transfer learning.

The surrounding rock classification methods mentioned above are either transformed into quantitative surrounding rock classification parameters after extracting the features from the images, and the final surrounding rock indexes are calculated for classification according to the corresponding relational formulas, or very time consuming, which is not practical in actual engineering. Among them, the transformation relationship between image features and surrounding rock classification parameters and the calculation formula of surrounding rock indexes are all obtained by the experience and expertise of those authors, whose accuracy and effectiveness are difficult to be guaranteed. At the same time, traditional DL models can be data-hungry and the performance relies heavily on the data sample size [22]. For complex engineering cases, it is therefore significant to investigate the deep transfer learning model based on EfficientNet with a compounding scaling technique. Further, scarcely did any former research studied the preprocessing of onsite tunnel face images and in this study, we contribute several techniques that facilitate the classification of surrounding rock with DL models.

This paper aims to develop a method that can quickly, objectively and accurately determine the surrounding rock grade in the construction stage. We try to combine a deep CNN with image processing techniques for surrounding rock classification, and explore a suitable CNN architecture based on tunnel face images, which will facilitate grading surrounding rocks explicitly from onsite photos taken at the tunnel.

The images were collected in batches during the construction process by using a digital camera system, and then the database of surrounding rock of tunnel face has been established. The surrounding rock is classified according to the Specifications for Design and Highway Tunnels (JTG3370-2018) and the opinions of experts. Then, based on the CNN EfficientNet [23], we combined transfer learning, data augmentation and other technologies, to classify the data set, and to calculate the accuracy, specificity, sensitivity and F1-score for a fair model evaluation. Our result has been compared with other CNN architectures, such as ResNet [24] and InceptionNet [25], to verify the accuracy and reliability of the model. Finally, we provide a systematic evaluation of the application of the deep transfer learning framework to classify surrounding rocks in tunnels. The overall method is shown in Fig.1.

2 Method and data set

2.1 EfficientNet

The CNN adopted in this paper is the EfficientNet model proposed by Tan and Le [23]. With the constant improvement of computer hardware level, the neural network model is becoming more and more bigger and the depth is more and more deeper. It can surely improve accuracy which, however, is based on the cost of computational efficiency and time consumption. Model compression [26] is a common method to improve efficiency by reducing model size at the cost of accuracy. The EfficientNet model combines efficiency and accuracy. Based on the three dimensions of model scaling: network depth, width and resolution, the scaling method proposed in Ref. [23] is used to achieve the uniformly scales at all dimensions at the same time with a simple yet effective compound coefficient, as shown in Fig.2.

The baseline network architecture EfficentNet-B0 is derived from MNASNet [27] and obtained by neural architecture search. The specific architecture is shown in Tab.1. Using the compound scaling method, EfficientNet-B1 to B7 is available. In this paper, we chose the minimum EfficientNet-B0 as the baseline network due to the small data set, followed by three fully connected layers and a dropout layer to avoid overfitting.

2.2 Transfer learning

Transfer learning is defined as “a machine learning method that uses existing knowledge to solve problems in different but related fields” [28], aiming at transferring knowledge between related domains [29]. For CNNs, transfer learning is the successful application of the “knowledge” gained from training on specific data sets to new domains. The general process of transfer learning of CNN is as follows.

1) Train random initialization parameters in the CNN using large data sets from related fields (such as ImageNet) prior to a specific application.

2) Perform feature extraction on specific data using the pretrained CNN.

3) Use the extracted features to train the pretrained CNN for a specific application.

Compared with the conventional method to train networks directly on the target data set, Zeiler and Fergus [30] pre-trained the CNN on the ImageNet data set and then applied it to the image classification data sets Caltech-101 [31,32] and Caltech-256 [33]. The accuracy of the image classification is improved by 40%. However, the domains of transfer learning are relatively close, as both ImageNet and Caltech are object recognition databases. Moreover, Donahue et al. [34] adopted a strategy similar to Zeiler’s [30] but applied transfer learning to the distant domains by pre-training an ImageNet-based CNN, including domain adaptation, subclass recognition, scene recognition, and so on.

In addition to studying transfer learning of CNN in various domains, Sharif et al. [35] also investigated the effectiveness of features learning at different CNN layers. The high-level features of CNNs have proved better transfer learning ability than the low-level ones. Zhou et al. [36] pre-trained two CNNs of the same structure by using a large image classification database (i.e., ImageNet) and scene recognition database (i.e., Places), and verified the transfer learning effect on a series of image classification and scene recognition databases. The experimental results show that domain correlation has a certain influence on the transfer learning of CNNs, indicating that the networks pre-trained by ImageNet and Places have better transfer learning effects on the databases of their respective domains.

In conclusion, transfer learning of CNN cannot only solve the problem of insufficient training samples of CNN under small data set conditions, but also greatly reduce training costs. In our model, in order to speed up the training and improve the model accuracy, we used the transfer learning technique with the ImageNet data set.

2.3 Data set

2.3.1 Surrounding rock grade

The classification of tunnel surrounding rock was first proposed in the form of qualitative description in Europe in 1774 [37]. Since 1920s, researchers have conducted a lot of studies on the classification of surrounding rock for underground cavities, and have put forward many theories and methods of surrounding rock classification. The methods vary from each other due to the different purposes. For example, to investigate the construction feasibility of surrounding rock, there are surrounding rock blasting and drillability classification methods [38]. To confirm the project construction planning, construction design and construction stability, there are surrounding rock stability classification methods. At present, the methods based on surrounding rock stability are widely used in tunnel and underground engineering.

After 2004, China reformulated the highway tunnel surrounding rock grading standards in the Highway Tunnel Design Code (JTG D70-2004), and classified the surrounding rock from intact to broken into Grade 1 to 5 by qualitative and quantitative evaluation methods. In this stage, the highway tunnel surrounding rock classification is mainly based on the rock mass classification standard issued by the State Construction Commission in 1994. That is, the basic quality index BQ value of rock mass is calculated according to the hardness degree of rock and the structural integrity degree of rock mass for preliminary classification. The calculation formula is as follows:

(1)BQ=90+3R_{\mathrm{c}}+250K{}_{\mathrm{v}},

where R_{\mathrm{c}} is the uniaxial compressive strength of rock, K_{\mathrm{v}} is the integrity coefficient of rock mass.

The following restrictions should be followed when calculating BQ:

• if R_{\mathrm{c}}>90K_{\mathrm{v}}+30, then R_{\mathrm{c}}=90K_{\mathrm{v}}+30;

• if K_{\mathrm{v}}>0.04R_{\mathrm{c}}+0.4, then K_{\mathrm{v}}=0.04R_{\mathrm{c}}+0.4.

Subsequently, the final [BQ] determines the final class of surrounding rock can be obtained by factors such as groundwater, occurrence of soft structural plane of rock mass and initial stress state of surrounding rock:

(2)\left[BQ\right]=BQ-100\left(K_1+K_2+K_3\right),

where K_1 is the groundwater correction coefficient, K_2 is the occurrence correction coefficient of rock weak structural plane, K_3 is the initial stress correction coefficient of surrounding rock.

At present, China’s highway tunnel surrounding rock classification is still based on this method, and some relevant scholars are still improving the BQ classification method [39–41]. In this paper, we presented the classification of surrounding rock according to the Specifications for Design and Highway Tunnels (JTG3370-2018). The data set in this application is collected from Xingyi City-ring Expressway located in the plateau mountainous area of south-west Guizhou Province, which is in the landform type of the middle to low dissolution mountains. It is characterized by karst peaks, forest peaks and valleys. The overlying quaternary residual slope deposit is silty clay, and the underlying bedrock is dolomite of middle Triassic Yangliujing Formation. In construction, the most common Grade 3 to 5 rocks are the main research objects.

2.3.2 Data collection

The shooting environment in the tunnel is poor, resulting in less information obtained by the camera. To ensure the integrity of photos, the camera fills random pixels into the photos, leading to increased noise in the photos. A large amount of dust in the air will also affect the transmission of light. Because of the diffuse reflection of light, dust will form many pointed light sources, forming numerous gray spots in the photo, and the characteristics of the tunnel face may be covered. Therefore, in addition to the high-resolution camera, some high-power lighting equipment is needed to ensure that the lighting conditions in the tunnel meet the shooting requirements, and dust removal is carried out by manual cleaning and high-pressure water washing, etc.

In the process of tunnel face image collection, in addition to eliminating the influence of dim lighting conditions and high dust concentration in the tunnel, attention should also be paid to the influence of shooting method on the image quality. In the process of shooting, the lighting should be uniform and parallel, so that the tunnel face will receive uniform illumination without producing shadows. At the same time, make sure that the camera lens is perpendicular to the tunnel face. Place the camera on the tripod horizontally and shoot from 3 to 4 m away from the face. Different face shooting methods are adopted for tunnels with different excavation methods.

Eventually we collect 921 surrounding rock images. After remove mislabeled images and other data cleaning techniques, 689 images are finally selected from the original rock database. The pictures are classified with three labels, 165 images of Grade 3 surrounding rock, 347 images of Grade 4 and 177 images of Grade 5. Fig.3 is the data set structure and some pictures of typical surrounding rocks. From the data constituents, we can find that the three grades of surrounding rock are imbalanced, therefore we applied the data augmentation to the selected image data sets, such as rotate and stretch the image, to make them balanced. The augmented images only used for training, did not participate in model testing.

2.3.3 Data preprocess

Images from preliminary data cleaning are still not ready as neural network input. A batch of preprocessing methods is necessary to improve the image quality and thus improve the efficiency and accuracy of the deep transfer learning model [42]. The original image sizes varies from pixels to pixels. To have a direct understanding of the pixel size distribution of the entire image data set, we draw the probability density distribution of pixel size, as shown in Fig.4. The image size is roughly concentrated at pixels, considering the limitation of computer hardware resources, calculation time and efficiency, we uniformly scaled the image to pixels, and the linear interpolation is our image scaling method. To improve the image quality, the image standardization can be applied, e.g., mean decentralization, so as to obtain the generalization effect more easily after training. The image needs to be normalized before feeding into the model, converting the RGB image pixel values from 0–255 to 0–1. Normalization will not change the image information, but for subsequent CNN, it can avoid vanishing/exploding gradient problem caused by too large pixel value range and reduce the chance of overfitting.

The equation of standardization and normalization are as follows:

(3)x_{\text{normalized}}=\frac{x_i-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)}{\mathrm{m}\mathrm{a}\mathrm{x}\left(x\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)},

(4)x_{\text{standardized}}=\frac{x_i-\mu }{\sigma },

where x_i is the image pixel value at position i, and \mathrm{m}\mathrm{i}\mathrm{n}\left(x\right) and \mathrm{m}\mathrm{a}\mathrm{x}\left(x\right) denote the minimum and maximum pixel values, respectively.

In addition, the image denoising can reduce the noise in the digital image, improve the smoothness. The image graying, convert image from three-channel image into single-channel image, can reduce the model complexity without losing image information. The image histogram equalization, enhance image contrast in dynamic range, make the image clearer. The schematics of the above methods are shown in Fig.5, where Gaussian filtering with kernel size is used in denoising method. Considering the model complexity and the loss of image information, we only adopt denoising with Gaussian filter kernel for data preprocessing.

3 Experiments

In this paper, the CNN model was built based on Google’s open-source framework TensorFlow. The Core hardware of the development platform was configured as Intel Core i5-12400F, NVIDIA GTX 3060Ti graphics card and 16 G memory. The code was written in Python, version 3.8.

3.1 Results

We trained our EfficientNet model by the hyperparameters setting as follows: Adamax optimizer with 0.9 momentum; batch norm momentum 0.99; weight decay 10⁻⁵; initial learning rate 10⁻³ that decays by 0.8 every 25 epochs. Due to hardware limitations, the batch size is set as 16, and the image size was set as 300 pixels × 300 pixels. All hyperparameters are shown in Tab.2. In this paper, a 5-fold cross validation is used to evaluate the performance of surrounding rock classification for the proposed model. We set the proportions of training, testing and validation samples to be 80%, 10%, and 10%, respectively, in each cross-validation fold. The validation data set is used for hyperparameters correction during training. As a result, the accuracy shall be improved and the chance of over-fitting shall be reduced.

Fig.6–Fig.9 illustrate the convergence graph indicating the evolution of loss and accuracy with the training steps accumulating for each model on training and validation data sets (the result of ResNet101 is not ideal so that we ignore it). Our EfficientNet model has performed over all other models, with nearly full training accuracy. IRV2 also performs well, with training accuracy on par with EfficientNet. The performance of ResNet50 is poor, with only 80% training accuracy. However, Fig.10 indicates that the validation loss for IRV2 increases after 100 epochs, which is a sign of overfitting. This is due to too many parameters in this model. In contrast, our EfficientNet-B0 model maintained high precision and showed no sign of overfitting, which proved the superiority of our model. Although ResNet50 has five times more layers than EfficientNet-B0, its accuracy is much lower, further indicating that increasing network depth does not necessarily improve accuracy. On the contrary, it may lead to overfitting and other problems. In addition, the convergence rate of other models is also lower than that of EfficientNet-B0, we can say that our model almost crushes all other models.

3.2 Model evaluation

In classification tasks, Accuracy, Precision, Recall and F1-score are the essential metrics, often used to evaluate the model performance. Those evaluation index can be calculated in Eqs. (5)−(9) with the 4 basic parameters in Tab.3, i.e., True Negative (TN), True Positive (TP), False Negative (FN), and False Positive (FP).

(5)\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN},

(6)\text{Precision}=\frac{TP}{TP+FP},

(7)\text{Specificity}=\frac{TN}{TN+FP},

(8)\text{Sensitivity}=\frac{TP}{TP+FN},

(9)\text{F1}\text{-score}=\frac{2\times \text{Sensiticity}\times \text{Precision}}{\text{Sensitivity}+\text{Precision}},

where Accuracy is the ratio of images correctly classified as true, Precision is the ratio of images is correctly judged as positive in all positive labels (i.e., true positives), and F1-score is a comprehensive metric to balance sensitivity and precision.

The evaluation of each model calculated through the testing data set is listed in Tab.4. As shown in the table, our EfficientNet model has the highest accuracy 89.86% from the 5-fold cross validation. Moreover, IRV2, with similar training accuracy, could not predict exactly on the test data set, with an accuracy of only 79.71%, indicating that the generalization of this model is poor and there is a serious overfitting problem. The box distribution figure of 5-fold cross-validation evaluation index of EfficientNet shows in Fig.10, where the red line represents the median value, the green triangle represents the mean value, the circle represents the outlier and the upper and lower bar represents the maximum and minimum values. After removing outliers, the fluctuation range of cross-validation results of each index is small, indicating a good stability of the model. The average test accuracy is 87.82%, which also meets the judgment requirements in practical construction.

We take the best result in the 5-fold cross validation and draw its confusion matrix in the Fig.11. The prediction accuracy of each grade is 94%, 90%, and 86% respectively, and the average prediction accuracy is 90%. The prediction accuracy of the Grade 3 surrounding rock is the highest, with only one test image was misjudged as the Grade 4. All the misjudgment data only occur in the neighboring grade, and there is no such unreasonable situation as the misjudgment of Grade 5 surrounding rock into Grade 3. Similarly, the Receiver Operating Characteristic (ROC) curve of each model can be drawn. The closer the ROC curve is to the upper left corner, the higher the True Positive Rate (TPR) and the lower the False Positive Rate (FPR), i.e., the higher the Sensitivity, the lower the misjudgment rate, and the better the predictive performance. The area under the curve is called the area under the curve (AUC), which is used to provide a quantitative measure of the performance. As shown in Fig.12, EfficientNet has the best performance with an AUC of 0.97 and a curve close to the upper left. ResNet50 and IRV2 are close to each other, with AUC of 0.94 and 0.93, respectively. ResNet101 has an AUC of only 0.68, which is basically of no value. The ROC curve further proves the above conclusion.

To further understand the internal mechanism of the classification model, a typical image of surrounding rock at each grade is selected and the Gradient-weight Class Activation Map (Grad-CAM) [43] is drawn. All images are preprocessed using the same method and the same training hyperparameters are used for each model, as shown in Fig.13. EfficientNet model shows more attention to the detailed regions of the image during the training process, such as the damaged and broken parts of surrounding rock and the area with color differences. These details are the classification basis of the model. Just like artificial classification of the surrounding rock, the model based on image recognition is also classified according to the integrity and degree of fragmentation of the surrounding rock. The only difference is that the model can give quantitative classification results through specific numerical values and characteristics without any interference from subjective factors. As long as surrounding rock images are input, no matter when, where and in what operating environment, the results of the same model are fixed. However, the artificial classification is subjective to some extent. In particular, the on-site surrounding rock classification is affected by the construction environment, project scheduling and other aspects, so the result of different technicians may also be different. Other models, ResNet50 and IRV2, failed to capture the images details, and as the model was trained, degradation occurred and the area of concern gradually shifted.

4 Conclusions

The proposed model combines data augmentation, image preprocessing, transfer learning and other techniques, and utilizes the state-of-the-art CNN model EfficientNet to successfully introduce DL into the tunnel face surrounding rock classification field. The prediction accuracy of the model finally reaches 89.96%, which is enough to meet the needs of actual engineering projects. Our data set is collected from Xingyi City Ring Expressway in the plateau mountainous area of Guizhou Province, the model developed has been successfully applied to this project, which proved the feasibility of our model. Our research has brought a new direction to the tunnel surrounding rock classification, and improved the work efficiency in the actual engineering construction, and saved the cost of time, manpower and material resources. We also explore the mechanism of DL and enhance the interpretability of the model, which can also provide a basis for manual classification.

References

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

[2] SunYWangXTangX. Deep convolutional network cascade for facial point detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New Jersey: IEEE, 2013: 3476–3483

[3] Siruvuri S V, Budarapu P R, Paggi M. Influence of cracks on fracture strength and electric power losses in Silicon solar cells at high temperatures: deep machine learning and molecular dynamics approach. Applied Physics A, Materials Science & Processing, 2023, 129(6): 408

[4] Sharma S, Awasthi R, Sastry Y S, Budarapu P R. Physics-informed neural networks for estimating stress transfer mechanics in single lap joints. Journal of Zhejiang University—Science A, 2021, 22(8): 621–631

[5] Huang H, Li Q, Zhang D. Deep learning based image recognition for crack and leakage defects of metro shield tunnel. Tunnelling and Underground Space Technology, 2018, 77: 166–176

[6] Aksoy C O, Geniş M, Aldaş G U, Özacar V, Özer S C, Yılmaz Ö. A comparative study of the determination of rock mass deformation modulus by using different empirical approaches. Engineering Geology, 2012, 131: 19–28

[7] Goh A T C, Zhang W. Reliability assessment of stability of underground rock caverns. International Journal of Rock Mechanics and Mining Sciences, 2012, 55: 157–163

[8] Cha Y J, Choi W, Suh G, Mahmoudkhani S, Büyüköztürk O. Autonomous structural visual inspection using region-based deep learning for detecting multiple damage types. Computer-Aided Civil and Infrastructure Engineering, 2018, 33(9): 731–747

[9] Ran X, Xue L, Zhang Y, Liu Z, Sang X, He J. Rock classification from field image patches analyzed using a deep convolutional neural network. Mathematics, 2019, 7(8): 755–771

[10] Chen J, Yang T, Zhang D, Huang H, Tian Y. Deep learning based classification of rock structure of tunnel face. Geoscience Frontiers, 2021, 12(1): 395–404

[11] Sharma A, Liu X, Yang X, Shi D. A patch-based convolutional neural network for remote sensing image classification. Neural Networks, 2017, 95: 19–28

[12] Nogueira K, Penatti O A B, Dos Santos J A. Towards better exploiting convolutional neural networks for remote sensing scene classification. Pattern Recognition, 2017, 61: 539–556

[13] Perol T, Gharbi M, Denolle M. Convolutional neural network for earthquake detection and location. Science Advances, 2018, 4(2): e1700578

[14] Patel A K, Chatterjee S. Computer vision-based limestone rock-type classification using probabilistic neural network. Geoscience Frontiers, 2016, 7(1): 53–60

[15] KwokC Y TWongM SHoH CLoF L CKoF W Y. Deep learning approach for rock outcrops identification. In: Proceedings of 2018 Fifth International Workshop on Earth Observation and Remote Sensing Applications (EORSA). New Jersey: IEEE, 2018: 1–6

[16] Alférez G H, Vázquez E L, Ardila A M M, Clausen B L. Automatic classification of plutonic rocks with deep learning. Applied Computing and Geosciences, 2021, 10: 100061

[17] ChengGGuoW. Rock images classification by using deep convolution neural network. Journal of Physics: Conference Series, 2017, 887(1): 012089

[18] dos Anjos C E M, Avila M R V, Vasconcelos A G P, Pereira Neta A M, Medeiros L C, Evsukoff A G, Surmas R, Landau L. Deep learning for lithological classification of carbonate rock micro-CT images. Computational Geosciences, 2021, 25(3): 971–983

[19] Liang Y, Cui Q, Luo X, Xie Z. Research on classification of fine-grained rock images based on deep learning. Computational Intelligence and Neuroscience, 2021, (1): 5779740

[20] Li J, Zhang L, Wu Z, Ling Z, Cao X, Guo K, Yan F. Autonomous Martian rock image classification based on transfer deep learning methods. Earth Science Informatics, 2020, 13(3): 951–963

[21] SimonyanKZissermanA. Very deep convolutional networks for large-scale image recognition. 2014, arXiv:1409.1556

[22] Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, Liu T, Wang X, Wang G, Cai J, Chen T. Recent advances in convolutional neural networks. Pattern Recognition, 2018, 77: 354–377

[23] TanMLeQ. Efficientnet: Rethinking model scaling for convolutional neural networks. 2019 arXiv: 1905.11946

[24] HeKZhangXRenSSunJ. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. New Jersey: IEEE, 2016: 770–778

[25] Szegedy C, Ioffe S, Vanhoucke V, Alemi A. Inception-v4, inception-resnet and the impact of residual connections on learning. Proceedings of the AAAI Conference on Artificial Intelligence, 2017, 31(1): 4278–4284

[26] HeYLinJLiuZWangHLiL JHanS. Amc: Automl for model compression and acceleration on mobile devices. In: Proceedings of the European Conference on Computer Vision (ECCV). Berlin: Springer, 2018: 784–800

[27] TanMChenBPangRVasudevanVSandlerMHowardALeQ V. Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Jersey: IEEE, 2019: 2820–2828

[28] Weiss K, Khoshgoftaar T M, Wang D D. A survey of transfer learning. Journal of Big Data, 2016, 3(1): 1–40

[29] Pan S J, Yang Q. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345–1359

[30] ZeilerM DFergusR. Visualizing and understanding convolutional networks. In: Proceedings of the European Conference on Computer Vision (ECCV). Berlin: Springer, 2014: 818–833

[31] Fei-Fei L, Fergus R, Perona P. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006, 28(4): 594–611

[32] KinnunenTKamarainenJ KLensuLLankinenJKälviäinenH. Making visual object categorization more challenging: Randomized caltech-101 data set. In: Proceedings of 20th International Conference on Pattern Recognition. New Jersey: IEEE, 2010: 476–479

[33] GriffinGHolubAPeronaP. Caltech-256 Object Category Dataset. Pasadena: California Institute of Technology, 2007

[34] DonahueJJiaYVinyalsOHoffmanJZhangNTzengEDarrellT. Decaf: A deep convolutional activation feature for generic visual recognition. In: Proceedings of the 31st International Conference on Machine Learning. Beijing: ACM, 2014: 647–655

[35] SharifR AAzizpourHSullivanJCarlssonS. CNN features off-the-shelf: an astounding baseline for recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. New Jersey: IEEE, 2014: 806–813

[36] ZhouBLapedrizaAXiaoJTorralbaAOlivaA. Learning deep features for scene recognition using places database. In: Advances in Neural Information Processing Systems. Massachusetts: MIT Press, 2014, 487–495

[37] Folk R L, Andrews P B, Lewis D W. Detrital sedimentary rock classification and nomenclature for use in New Zealand. New Zealand Journal of Geology and Geophysics, 1970, 13(4): 937–968

[38] White C G. A rock drillability index. Rocks & Minerals, 1969, 44(7): 490–490

[39] Shen Y J, Yan R X, Yang G S, Xu G L, Wang S Y. Comparisons of evaluation factors and application effects of the new [BQ] GSI system with international rock mass classification systems. Geotechnical and Geological Engineering, 2017, 35(6): 2523–2548

[40] Jefferies M G, Davies M P. Soil classification by the cone penetration test. Canadian Geotechnical Journal, 1991, 28(1): 173–176

[41] Guo S, Qi S, Saroglou C A B Q. a classification system for anisotropic rock mass based on China National Standard. Journal of Central South University, 2020, 27(10): 3090–3102

[42] Verma H, Siruvuri S V, Budarapu P R. A machine learning-based image classification of silicon solar cells. International Journal of Hydromechatronics., 2024, 7(1): 49–66

[43] ZhouBKhoslaALapedrizaAOlivaATorralbaA. Learning deep features for discriminative localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New Jersey: IEEE, 2016: 2921–2929

RIGHTS & PERMISSIONS

Higher Education Press