Rock microcrack identification and detection method based on deep learning

Abstract

A deep learning-based rock microcrack identification and detection method belongs to the technical field of rock microcrack detection and identification in geological resources and geological engineering, and greatly improves the detection precision of rock microcracks by integrating a multi-scale feature pre-extraction module (MFP), an encoder space cascade coding path (ESCE) and a decoder input SE module. The conditions of false detection and missing detection are reduced, and the precision and the stability of micro-crack detection in the rock are remarkably improved. According to the method, through a reasonably designed network structure, the calculation amount of the network is reduced, the detection process is made to be more lightweight, and the method is suitable for rapid processing and real-time application of a large-scale data set. According to the method, the requirement for manual participation is greatly reduced, automatic detection and classification of the microcracks in the rock are achieved through the deep learning algorithm, the reliability and efficiency of rock integrity and mechanical property evaluation are improved, and the method has wide geological engineering application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of rock microcrack detection and identification in geological resources and geological engineering, and in particular relates to a rock microcrack identification and detection method based on deep learning. Background Art

[0002] Rocks usually contain microcracks of varying degrees that are formed by natural conditions or artificial induction. On the one hand, microcracks can significantly affect the deformation characteristics of rocks and may endanger the stability of rock or rock mass structures. If these cracks can be fully characterized, potential structural damage can be reduced and the probability of geological disasters can be reduced. On the other hand, microcracks can weaken rock strength, reduce the difficulty of excavation and drilling, and improve the efficiency of rock fracture and crushing. Therefore, accurate identification and characterization of microcracks are crucial for accurately predicting and evaluating rock properties, and are of great significance to the field of geological resources and geological engineering in disaster prevention and mitigation and promoting the efficient development and utilization of energy resources.

[0003] Microcracks in rocks are characterized by large numbers and tiny structures. The existing technology has problems such as incomplete extraction of microcrack image information and incomplete characterization of feature details, resulting in unsatisfactory fine quantification of microcrack parameters. Although some conventional detection methods such as optical microscopes can provide visualization effects in crack detection, complex microcrack morphologies often require further fine identification to obtain accurate statistical parameters, and this technology is costly, time-consuming and technically demanding. In addition, for cracks that have been detected and identified, the general method is to binarize the crack image in RGB format and then extract the crack parameters through skeletonization. The biggest problem with this method is that it will lose a large amount of real details and information of microcracks, resulting in inaccurate microcrack identification detection and subsequent statistical quantification. Therefore, realizing automated, intelligent and high-reliability identification and detection of microcracks in rocks is an urgent problem to be solved in rock performance evaluation.

[0004] The rapid development of artificial intelligence technology has provided a new solution for the automated detection and accurate identification of microcracks in rocks. The convolutional neural network target detection algorithm has the advantages of fast and accurate recognition and has been widely used in object and image recognition. However, due to the small and large number of microcracks in rocks, the minerals are intertwined and have different colors and lusters, and microcracks are often blurred and obscured. In addition, due to the complexity of the detection environment, there is a high misjudgment rate for the identification and extraction of microcracks.

[0005] Therefore, the prior art urgently needs a new technical solution to solve the above problems. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a rock microcrack identification and detection method based on deep learning. The method adopts an improved network structure based on U-Net, called MES-Net. Through the multi-scale feature extraction of the MES-Net network, the method can accurately locate and extract the detailed features of microcracks while performing efficient training, thereby providing a reliable basis and reference for the accurate prediction and evaluation of rock properties.

[0007] A rock microcrack identification and detection method based on deep learning includes the following steps: Step 1: Visualize the microcracks in the target rock by vacuum impregnation with fluorescent resin to obtain a fluorescent image in RGB format; Step 2: Perform data enhancement processing on the fluorescence image obtained in step 1 to establish a deep learning database; Step 3: Establish a U-Net network model, including a set of downsampling encoding paths and upsampling decoding paths. The downsampling encoding path consists of several 3×3 convolutional layers Conv and 2×2 maximum pooling layers Max Pooling. After each convolutional layer, a ReLU (Rectified Linear Unit, linear rectification function) activation function is used to increase the nonlinear expression ability of the network; the upsampling decoding path uses a series of 2×2 upsampling Up-sampling operations to gradually increase the spatial resolution of the feature map, so that the feature map is restored to its original size; Step 4: Establish a multi-scale feature pre-extraction MFP module to pre-extract features from input images of different scales and obtain a spliced ​​feature map; Step 5: Use the concatenated feature map as the input of the encoder spatial cascade coding path ESCE. The encoder spatial cascade coding path ESCE consists of three consecutive 3×3 dilated convolution operations with dilation rates of 1, 3, and 5 respectively. The feature maps obtained after each layer of convolution are concatenated together as the output feature map; the output feature map is concatenated with the output feature map of the fourth layer encoder as the input of the decoder; Step 6: Insert the decoder input module SE into the decoder, perform global average pooling, process with two fully connected layers FC Laye, and fit the Sigmoid activation function to obtain the channel weighted vector s, and multiply it with the original feature map of the corresponding channel to generate the output feature map; Step 7: Divide the data set in the database of step 2 into training set and test set in a ratio of 8:2. Input the test data after model training into the MES-Net model trained in steps 3 to 6 for prediction. Extract the microcrack detection results through the grayscale image output by the model.

[0008] The method for obtaining the spliced ​​feature map described in step 4 is: first, continuously downsample the input image to obtain four images of different resolutions; then, use the MFP module to extract semantic features from these images; finally, splice the obtained feature map with the input feature map of the corresponding encoding layer in U-Net.

[0009] The model training process in step seven is as follows: use the cross-entropy loss function to measure the difference between the predicted result and the actual label, and use the Adam optimizer to optimize the model; use the early stopping mechanism to prevent overfitting, and monitor the validation set loss value of each epoch. Epoch is the number of training rounds, which means the training process of traversing the entire training data set once; when the validation set loss value does not decrease significantly in several consecutive epochs, stop training early.

[0010] The multi-scale feature pre-extraction MFP module described in step 4 includes 5 branches. The first three branches from top to bottom are composed of two 3×3 convolution kernels with dilation rates of 1 and 3, and a 5×5 convolution kernel with dilation rate of 1; the fourth branch connects a 3×3 standard convolution kernel with a 1×1 convolution kernel; the last branch uses two cascaded 3×3 convolution kernels with dilation rates of 3 and 5 respectively.

[0011] Through the above design scheme, the present invention can bring the following beneficial effects: A rock microcrack identification and detection method based on deep learning, by integrating a multi-scale feature pre-extraction module (MFP), an encoder spatial cascade coding path (ESCE) and a decoder input SE module, greatly improves the detection accuracy of rock microcracks, reduces false detections and missed detections, and significantly improves the accuracy and stability of microcrack detection in rocks. This method reduces the amount of network calculations through a reasonably designed network structure, making the detection process more lightweight and suitable for rapid processing and real-time application of large-scale data sets. At the same time, the present invention greatly reduces the need for manual participation, realizes the automated detection and classification of microcracks in rocks through a deep learning algorithm, improves the reliability and efficiency of rock integrity and mechanical property evaluation, and has broad application prospects in geological engineering.

[0012] Furthermore, the present invention realizes multi-scale semantic feature extraction of images with different resolutions through the MFP module, and deeply processes the feature map through the ESCE path, ensuring the accurate capture and reconstruction of microcrack details in the rock. The present invention introduces the SE module in the decoding path, effectively weakens the semantic difference between the encoding path and the decoding path, improves the weighted selectivity of the feature map channel, and enhances the model's attention to significant features. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention is further described below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a schematic diagram of the network structure of a rock microcrack identification and detection method based on deep learning in the present invention.

[0014] Figure 2 This is a schematic diagram of the MFP module structure of a rock microcrack identification and detection method based on deep learning in the present invention.

[0015] Figure 3 This is a schematic diagram of the ESCE path structure of a rock microcrack identification and detection method based on deep learning in the present invention.

[0016] Figure 4 This is a schematic diagram of the SE module structure of a rock microcrack identification and detection method based on deep learning in the present invention.

[0017] Figure 1 middle: Input Image: Input image refers to the original image used for model processing; Output Image: Output image, which refers to the result after model processing; Skip Connection: Skip connection, which directly concatenates the feature map in the downsampling path with the feature map in the upsampling path; MFP: Multi-scale Feature Pre-extraction, a multi-scale feature pre-extraction module, is used to extract multi-scale semantic features from input images of different resolutions; SE: Squeeze-and-Excitation, decoder input module, used to reduce the semantic difference between encoding and decoding paths; ESCE Path: Encoder Spatial Cascading Encoding Path, encoder spatial cascading encoding path, which improves the feature extraction depth and enhances the feature reuse rate by connecting multiple convolutional layers in series; Conv: convolution layer, 3×3, 1×1, etc. are the sizes of the convolution kernels in the convolution layer; ReLu: Rectified Linear Unit, linear rectification activation function; Sigmoid: Sigmoid activation function is used later; Max Pool 2×2: 2×2 maximum pooling layer, which reduces the size of feature maps and is used for downsampling feature maps; Up-Conv 2×2: 2×2 upsampling convolution layer, which increases the size of the feature map and is used for upsampling feature maps; Feature maps: Feature maps are the output of the convolutional layer; Concatenated feature maps: The process of concatenating multiple feature maps together.

[0018] Figure 2 middle: Input Feature Maps: Input feature maps; Output Feature Maps: Output feature maps; Rate: dilation rate, which controls the coverage of the convolution kernel and can extract multi-scale features; BN: Batch Normalization, batch normalization; Figure 3 middle: Channel: The number of channels, which refers to the depth of the feature map.

[0019] Figure 4 middle: Encodr Feature Maps: Encoder feature maps, which refer to the feature maps output by the network encoder part; Decodr Upsampling Feature Maps: Decoder upsampling feature maps, which refers to the feature maps obtained by upsampling in the network decoder part; Concatenate: concatenates feature maps from different sources in the channel dimension; Multiply: Multiply the channel weights of the Sigmoid output by the original feature map element by element to achieve feature recalibration; Gloabla pooling: global pooling; FC Layer: fully connected layer. DETAILED DESCRIPTION

[0020] A rock microcrack identification and detection method based on deep learning, such as Figure 1-Figure 4As shown in the figure, this method adopts an improved network structure based on U-Net, called MES-Net, which includes three key modules: Multi-scale Feature Pre-extraction (MFP), Encoder Spatial Cascading Encoding (ESCE), and Decoder Input Module (Squeeze-and-Excitation, SE). The MFP module is used to extract multi-scale semantic features from input images of different resolutions. The ESCE path improves the feature extraction depth and enhances the feature reuse rate by connecting multiple convolutional layers in series. The SE module is used to reduce the semantic difference between the encoding path and the decoding path. Through the multi-scale feature extraction of the MES-Net network, this method can accurately locate and extract the detailed features of microcracks while efficiently training, providing a reliable basis and reference for the accurate prediction and evaluation of rock properties.

[0021] The specific method is as follows: Step 1: Use fluorescent resin vacuum impregnation to visualize the microcracks in the target rock. The specific steps are as follows: First, mix the fluorescent dye with the two-component epoxy resin and impregnate the rock sample under vacuum conditions for 2 hours. Then, transfer the impregnated rock sample to a 25°C constant temperature box for curing for at least 9 hours. After that, use 365nm ultraviolet light to illuminate the key area to be observed and take a fluorescent image of the microcracks.

[0022] Step 2: Perform data enhancement processing on the acquired RGB format fluorescence image. Data enhancement techniques include scaling, cropping, brightness adjustment, contrast adjustment, hue adjustment, saturation adjustment, Gaussian blur, noise addition, etc. These data enhancement techniques can be used alone or in combination to generate richer training data and establish a deep learning database, thereby improving the generalization ability of the model and its adaptability to real scenes.

[0023] Step 3: Design of U-Net basic model. The MES-Net model is based on the U-Net architecture. The U-Net model consists of a set of downsampling encoding paths and upsampling decoding paths. The encoding path is composed of multiple 3×3 convolutional layers (Conv) and 2×2 maximum pooling layers (Max Pooling). After each convolutional layer, the ReLU (Rectified Linear Unit) activation function is used to increase the nonlinear expression ability of the network. Each pooling operation will halve the spatial resolution of the feature map and double the number of feature channels, thereby extracting more abstract and deeper semantic information.

[0024] The decoding path uses a series of 2×2 up-sampling operations to gradually increase the spatial resolution of the feature map and restore the feature map to its original size. After each up-sampling operation, it is concatenated with the feature map of the corresponding layer in the encoding path to ensure the fusion of global and local features.

[0025] Through skip connections, the detailed semantic features of the shallow layer can be fused with the abstract semantic features of the deep layer, allowing the network to balance features of different sizes and depths. Finally, the deep feature map from the last decoding layer undergoes 3×3 convolution and sigmoid activation. The number of channels of the network output can be adjusted according to the segmentation target. Each channel represents the pixel probability map of the corresponding segmentation target, which can achieve effective fusion of multi-scale features.

[0026] Step 4: The MFP module is designed to pre-extract features from input images of different scales. First, the input image is continuously downsampled to obtain four images of different resolutions. Then, the MFP module is used to extract semantic features from these images. Finally, the obtained feature map is concatenated with the input feature map of the corresponding encoding layer in U-Net. Each branch of the MFP module consists of convolution kernels of different sizes and dilation rates. Figure 2 As shown in the figure, the first three branches from top to bottom are composed of two 3×3 convolution kernels with dilation rates of 1 and 3, and a 5×5 convolution kernel with dilation rate of 1. The fourth branch connects a 3×3 standard convolution kernel with a 1×1 convolution kernel. In the last branch, two cascaded 3×3 convolution kernels with dilation rates of 3 and 5 are used. As mentioned above, the combination of U-Net and MFP modules enables the network to extract semantic features of different scales in advance, so that the network has the ability to detect both large and small objects.

[0027] Step 5: The ESCE path can avoid information loss and make full use of multi-scale features. Figure 1 As shown in , the input feature maps of the second and third encoder layers are mapped to the same feature space as the fourth encoder layer through the maximum pooling operation. These feature maps are then concatenated as the input of the ESCE path. The ESCE path consists of three consecutive 3×3 dilated convolution operations with dilation rates of 1, 3, and 5, respectively, as shown in Figure 3 As shown in Figure 2. The feature maps obtained after each convolution layer are concatenated together as the output feature map. Finally, the output feature map of the ESCE path is concatenated with the output feature map of the fourth layer encoder as the input of the decoder.

[0028] In step 6, the decoder input SE module is built. In the original U-Net model, the encoding path and the decoding path are directly connected through skip connections, but there is a semantic gap between them. In order to reduce this gap and automatically adjust the contribution of different feature maps, the SE module is inserted at the input of the decoder. The SE module obtains a 1×1×L one-dimensional feature vector z from the decoder input feature map through a global average pooling operation. Subsequently, z is processed by two fully connected layers (FC Layer). The number of neurons in the first FC Layer is 1×1×(L / 16), the activation function is ReLU, and the number of neurons in the second FC layer is 1×1×L. Finally, the complex nonlinear relationship between different channels is fitted by the Sigmoid activation function to generate a channel weighted vector s of size 1×1×L, which is multiplied with the original feature map of the corresponding channel to generate the output feature map. Through this method, the decoder path can pay more attention to the salient feature map channels and reduce the interference of useless channels.

[0029] In step 7, microcrack target recognition and detection. The data set that has been processed by data enhancement in step 2 is divided into a training set and a test set in a ratio of 8:2. During the training process, the cross-entropy loss function is used to measure the difference between the predicted result and the actual label, and the Adam optimizer is used to optimize the model. In order to prevent overfitting, the early stopping mechanism is adopted, and the validation set loss value of each epoch is monitored. When the validation set loss value does not decrease significantly in several consecutive epochs, the training is stopped in advance. After the training is completed, the test data is input into the trained MES-Net model for prediction. The detection results of microcracks are extracted through the grayscale image output by the model. To further improve the detection accuracy, post-processing techniques can be used to refine the detection results, such as removing noise and microcracks with broken connections through morphological operations, and using threshold segmentation techniques to binarize the grayscale image to extract clear microcrack contours. Ultimately, the microcrack detection method based on the MES-Net model can generate accurate microcrack identification images, providing an important basis for subsequent rock performance analysis and geological disaster early warning.

[0030] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

[0031] For example, without changing the basic architecture of MES-Net, similar detection effects can be achieved by adjusting the convolution kernel size and dilation rate in the MFP module, or by using different activation functions or normalization techniques in the SE module to optimize the model performance.

Claims

1. A rock microcrack identification and detection method based on deep learning, characterized by: The following steps are included: Step 1: Visualize the microcracks in the target rock by vacuum impregnation with fluorescent resin to obtain a fluorescent image in RGB format; Step 2: Perform data enhancement processing on the fluorescence image obtained in step 1 to establish a deep learning database; Step 3: Establish a U-Net network model, including a set of downsampling encoding paths and upsampling decoding paths. The downsampling encoding path consists of several 3×3 convolutional layers Conv and 2×2 maximum pooling layers Max Pooling. After each convolutional layer, the ReLU linear rectification activation function is used to increase the nonlinear expression ability of the network; the upsampling decoding path uses a series of 2×2 upsampling Up-sampling operations to gradually increase the spatial resolution of the feature map, so that the feature map is restored to its original size; Step 4: Establish a multi-scale feature pre-extraction MFP module to pre-extract features from input images of different scales and obtain a spliced ​​feature map; Step 5: Use the concatenated feature map as the input of the encoder spatial cascade coding path ESCE. The encoder spatial cascade coding path ESCE consists of three consecutive 3×3 dilated convolution operations with dilation rates of 1, 3, and 5 respectively. The feature maps obtained after each layer of convolution are concatenated together as the output feature map; the output feature map is concatenated with the output feature map of the fourth layer encoder as the input of the decoder; Step 6: Insert the decoder input module SE into the decoder, perform global average pooling, process with two fully connected layers FCLeye, and fit the Sigmoid activation function to obtain the channel weighted vector s, and multiply it with the original feature map of the corresponding channel to generate the output feature map; Step 7: Divide the data set in the database of step 2 into training set and test set in a ratio of 8:

2. Input the test data after model training into the MES-Net model trained in steps 3 to 6 for prediction. Extract the microcrack detection results through the grayscale image output by the model.

2. According to the method for identifying and detecting rock microcracks based on deep learning in claim 1, it is characterized by: step The method for obtaining the spliced ​​feature map is as follows: first, the input image is continuously downsampled to obtain four images with different resolutions; then, the MFP module is used to extract semantic features from these images; finally, the obtained feature map is spliced ​​with the input feature map of the corresponding encoding layer in U-Net.

3. The rock microcrack identification and detection method based on deep learning according to claim 1 is characterized in that: The model training process in step seven is as follows: use the cross-entropy loss function to measure the difference between the predicted result and the actual label, and use the Adam optimizer to optimize the model; use the early stopping mechanism EarlyStopping to prevent overfitting, and monitor the validation set loss value of each epoch training round; when the validation set loss value does not decrease significantly in several consecutive epochs, stop training in advance.

4. According to the method for identifying and detecting rock microcracks based on deep learning in claim 1, it is characterized by: step The multi-scale feature pre-extraction MFP module includes 5 branches. The first three branches from top to bottom are composed of two 3×3 convolution kernels with dilation rates of 1 and 3, and a 5×5 convolution kernel with dilation rate of 1; the fourth branch connects a 3×3 standard convolution kernel with a 1×1 convolution kernel; the last branch uses two cascaded 3×3 convolution kernels with dilation rates of 3 and 5 respectively.

Images