A rock micro-crack identification and detection method based on deep learning

By using the MES-Net network structure, combined with multi-scale feature extraction and deep learning, the problems of accuracy and efficiency in rock microcrack detection are solved, enabling reliable evaluation and efficient detection of rock properties.

CN119693794BActive Publication Date: 2025-11-07JILIN UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411749433.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-12-02
Publication Date
2025-11-07
Estimated Expiration
2044-12-02

AI Technical Summary

Technical Problem

Existing technologies are insufficient to accurately identify and characterize microcracks in rocks, resulting in incomplete detection and incomplete characterization of features. Furthermore, conventional methods are costly, time-consuming, and have a high false positive rate.

Method used

An improved network structure, MES-Net, is adopted, which combines a multi-scale feature extraction module (MFP), an encoder spatial cascaded encoding path (ESCE), and a decoder input module (SE). Through deep learning, the automatic identification of rock microcracks is achieved. A high-efficiency deep learning database is established by using fluorescent resin vacuum impregnation and data augmentation processing.

Benefits of technology

It significantly improves the accuracy and stability of rock microcrack detection, reduces false detections and missed detections, lowers the computational load, and achieves reliability and efficiency in rock performance evaluation. It is suitable for rapid processing and real-time application of large-scale datasets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119693794B_ABST
    Figure CN119693794B_ABST
Patent Text Reader

Abstract

A rock micro crack recognition detection method based on deep learning belongs to the technical field of rock micro crack detection and recognition in geological resources and geological engineering, through integration of a multi-scale feature pre-extraction module (MFP), an encoder space cascade coding path (ESCE) and a decoder input SE module, the detection precision of rock micro cracks is greatly improved, the false detection and missed detection conditions are reduced, and the precision and stability of rock micro crack detection are significantly improved. Through the reasonable design of the network structure, the calculation amount of the network is reduced, so that the detection process is more lightweight, and it is suitable for large-scale data set fast processing and real-time application. The application greatly reduces the demand for manual participation, realizes the automatic detection and classification of rock micro cracks through the deep learning algorithm, improves the reliability and efficiency of rock integrity and mechanical property evaluation, and has wide geological engineering application prospect.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of rock micro-crack detection and identification in geological resources and geological engineering, and particularly relates to a rock micro-crack identification and detection method based on deep learning. BACKGROUND

[0002] Rock usually contains micro-cracks of different degrees formed by natural conditions or human-induced. On the one hand, micro-cracks can significantly affect the deformation characteristics of rock, which may endanger the stability of rock or rock mass structure. If these cracks can be fully characterized, the potential structural damage can be reduced, and the probability of geological disasters can be reduced. On the other hand, micro-cracks can weaken the strength of rock, reduce the difficulty of excavation and drilling, and improve the efficiency of rock rupture and fragmentation. Therefore, accurate identification and characterization of micro-cracks are crucial for accurate prediction and evaluation of rock performance, and are of great significance to disaster prevention and mitigation and efficient development and utilization of energy resources in the field of geological resources and geological engineering.

[0003] Micro-cracks in rock have the characteristics of large quantity and small structure. The existing technology has the problems of incomplete extraction of micro-crack image information and incomplete characterization of feature details, resulting in unsatisfactory fine quantization of micro-crack parameters. Some conventional detection methods such as optical microscope can provide visualization effect in crack detection, but complex micro-crack morphology often needs further fine identification to obtain accurate statistical parameters. Moreover, this technology has high cost, long time consumption and high technical requirements. In addition, for the cracks that have been detected and identified, the general method is to perform binaryzation processing on the crack image in RGB format, and then realize the extraction of crack parameters through skeletonization. The biggest problem of this method is that a large amount of real details and information of micro-cracks will be lost, resulting in inaccurate micro-crack identification and detection and subsequent statistical quantization. Therefore, it is an urgent problem to be solved in rock performance evaluation to realize automatic and intelligent high-reliability identification and detection of micro-cracks in rock.

[0004] The rapid development of artificial intelligence technology provides a new solution for automatic detection and accurate identification of micro-cracks in rock. Convolutional neural network object detection algorithm has the advantages of fast and accurate identification, and has been widely used in object and image recognition. However, due to the small size and large quantity of micro-cracks in rock, the minerals are mixed with each other and have different colors and luster, and the micro-cracks often appear blurred and shielded, plus the complexity of the detection environment, the identification and extraction of micro-cracks are prone to high misjudgment rate.

[0005] Therefore, the existing technology urgently needs a new technical solution to solve the above problems. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a rock micro-crack identification and detection method based on deep learning, which adopts an improved network structure based on U-Net, called MES-Net, and through multi-scale feature extraction of the MES-Net network, the method can accurately locate and extract the detailed features of micro-cracks while being trained efficiently, providing reliable basis and reference for accurate prediction and evaluation of rock performance.

[0007] A rock micro-crack identification and detection method based on deep learning, comprising the following steps,

[0008] Step one, adopt fluorescent resin vacuum impregnation to visualize the micro-cracks in the target rock, and obtain a fluorescent image in RGB format;

[0009] Step two, perform data enhancement processing on the fluorescent image obtained in step one, and establish a deep learning database;

[0010] Step three, establish a U-Net network model, including a set of down-sampling encoding path and up-sampling decoding path, the down-sampling encoding path includes a plurality of 3x3 convolution layers Conv and 2x2 maximum pooling layers Max Pooling, after each convolution layer, a ReLU (Rectified Linear Unit) activation function is used to increase the non-linear expression ability of the network; the up-sampling decoding path uses a series of 2x2 up-sampling operations to gradually increase the spatial resolution of the feature map, so that the feature map is restored to the original size;

[0011] Step four, establish a multi-scale feature pre-extraction MFP module for pre-extracting features in input images of different scales and obtaining a spliced feature map;

[0012] Step five, splice the feature map as the input of the encoder spatial cascade encoding path ESCE, the encoder spatial cascade encoding path ESCE is composed of three consecutive 3x3 dilated convolution operations with dilated rates of 1, 3 and 5, the feature maps obtained after each layer of convolution are spliced together as the output feature map; the output feature map is spliced with the output feature map of the fourth layer of encoder as the input of the decoder;

[0013] Step six, insert a decoder input module SE in the decoder, perform processing through a global average pooling operation, two fully connected layers FC Laye, and a Sigmoid activation function fitting, obtain a channel weighting vector s, and multiply it with the original feature map of the corresponding channel to generate an output feature map;

[0014] Step seven, the data set in the database of step two is divided into a training set and a test set in a ratio of 8:2, the test data after model training is input into the MES-Net model trained in steps three to six for prediction, and the detection result of micro cracks is extracted through the gray image output by the model.

[0015] The method for obtaining the spliced feature map in step four is: first, continuously down-sampling the input image to obtain four images with different resolutions; then, using the MFP module to extract semantic features from these images; finally, splicing the obtained feature map with the input feature map of the corresponding encoding layer in the U-Net.

[0016] The model training process in step seven is: using the cross-entropy loss function Cross-Entropy Loss to measure the difference between the prediction result and the actual label, and using the Adam optimizer Adam Optimizer to optimize the model; using the early stopping mechanism Early Stopping to prevent overfitting, and monitoring the validation set loss value of each epoch, epoch is the number of training rounds, indicating the training process of traversing the entire training data set once; when the validation set loss value does not significantly decrease in consecutive epochs, the training is stopped in advance.

[0017] The multi-scale feature pre-extraction MFP module in step four includes 5 branches, the first three branches from top to bottom are composed of two 3*3 convolution kernels with expansion rates of 1 and 3, and a 5*5 convolution kernel with an expansion rate of 1; the fourth branch connects a 3*3 standard convolution kernel with a 1*1 convolution kernel; and the last branch adopts two cascaded 3*3 convolution kernels with expansion rates of 3 and 5.

[0018] Through the above design scheme, the present application can bring the following beneficial effects: a rock micro crack recognition and detection method based on deep learning, by integrating a multi-scale feature pre-extraction module (MFP), an encoder spatial cascade encoding path (ESCE) and a decoder input SE module, the detection accuracy of rock micro cracks is greatly improved, the false detection and missed detection are reduced, and the accuracy and stability of rock micro crack detection are significantly improved. Through the reasonable design of the network structure, the calculation amount of the network is reduced, so that the detection process is more lightweight, and it is suitable for large-scale data set fast processing and real-time application. At the same time, the present application greatly reduces the demand for manual participation, realizes the automatic detection and classification of micro cracks in rock through deep learning algorithm, improves the reliability and efficiency of rock integrity and mechanical property evaluation, and has wide geological engineering application prospect.

[0019] Further, the present application realizes multi-scale semantic feature extraction of different resolution images through the MFP module, and performs deep processing on the feature map through the ESCE path, thereby ensuring accurate capture and reconstruction of micro-crack details in the rock. BRIEF DESCRIPTION OF DRAWINGS

[0020] The present application will be further described below in conjunction with the drawings and specific embodiments:

[0021] Figure 1 A network structure diagram of the rock micro-crack recognition and detection method based on deep learning.

[0022] Figure 2 A MFP module structure diagram of the rock micro-crack recognition and detection method based on deep learning.

[0023] Figure 3 An ESCE path structure diagram of the rock micro-crack recognition and detection method based on deep learning.

[0024] Figure 4 An SE module structure diagram of the rock micro-crack recognition and detection method based on deep learning.

[0025] Figure 1 In the present application:

[0026] Input Image: input image, refers to the original image used for model processing;

[0027] Output Image: output image, refers to the result after model processing;

[0028] Skip Connection: skip connection, directly splices the feature map in the down-sampling path with the feature map in the up-sampling path;

[0029] MFP: Multi-scale Feature Pre-extraction, multi-scale feature pre-extraction module, used for extracting multi-scale semantic features from input images of different resolutions;

[0030] SE: Squeeze-and-Excitation, decoder input module, used for reducing the semantic difference between the encoding path and the decoding path;

[0031] ESCE Path: Encoder Spatial Cascading Encoding Path, an encoder spatial cascading encoding path that improves the depth of feature extraction and enhances the reuse rate of features by connecting multiple convolution layers in series;

[0032] Conv: Convolution layer, 3x3, 1x1, etc. are the sizes of the convolution kernel in the convolution layer;

[0033] ReLu: Rectified Linear Unit, a linear rectified activation function;

[0034] Sigmoid: Sigmoid activation function is used at the end;

[0035] Max Pool 2x2: 2x2 Max Pooling layer, reduces the size of the feature map, used for down-sampling the feature map;

[0036] Up-Conv 2x2: 2x2 Up-sampling Convolution layer, increases the size of the feature map, used for up-sampling the feature map;

[0037] Feature maps: Feature maps, the output of the convolution layer;

[0038] Concatenated feature maps: The process of concatenating multiple feature maps together.

[0039] Figure 2 In:

[0040] Input Feature Maps: Input feature maps;

[0041] Output Feature maps: Output feature maps;

[0042] Rate: Dilated rate, controls the coverage of the convolution kernel, can extract multi-scale features;

[0043] BN: Batch Normalization, batch normalization;

[0044] Figure 3 In:

[0045] Channel: Channel number, refers to the depth of the feature map.

[0046] Figure 4 In:

[0047] Encodr Feature Maps: Encoder feature maps, refers to the feature maps output by the network encoder part;

[0048] Decodr Upsampling Feature Maps: Decoder up-sampling feature maps, referring to the feature maps obtained by the network decoder part through up-sampling;

[0049] Concatenate: Concatenation, merging feature maps from different sources in the channel dimension;

[0050] Multiply: Multiplication, multiplying the channel weight of the Sigmoid output with the original feature map element by element to realize feature re-calibration;

[0051] Gloabla pooling: Global pooling;

[0052] FC Layer: Fully connected layer. DETAILED DESCRIPTION

[0053] A rock micro-crack recognition and detection method based on deep learning, as shown in Figures 1-4 The method uses an improved network structure based on U-Net, called MES-Net, which includes three key modules: multi-scale feature pre-extraction module (MFP), encoder spatial cascading encoding path (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 concatenating multiple convolution layers, and the SE module is used to reduce the semantic difference between the encoding path and the decoding path. Through multi-scale feature extraction of the MES-Net network, the method can efficiently train while accurately positioning and extracting the detailed features of micro-cracks, providing reliable basis and reference for accurate prediction and evaluation of rock performance.

[0054] The specific method is as follows:

[0055] Step 1, the micro-cracks in the target rock are visualized by vacuum impregnation of fluorescent resin. The specific steps are as follows: first, mix the fluorescent dye with the two-component epoxy resin, and immerse the rock sample under vacuum for 2 hours. Then, transfer the immersed rock sample to a constant temperature oven at 25°C and cure for at least 9 hours. After that, use 365nm ultraviolet light to irradiate the key area to be observed, and take the fluorescence image of the micro-cracks.

[0056] Step 2, data augmentation processing is performed on the acquired RGB format fluorescence image. Data augmentation techniques include scaling, cropping, brightness adjustment, contrast adjustment, hue adjustment, saturation adjustment, Gaussian blur, noise addition, etc. These data augmentation techniques can be used alone or in combination to generate more abundant training data, establish a deep learning database, and thus improve the generalization ability of the model and its adaptability to real scenes.

[0057] Step 3, U-Net base model design. The MES-Net model is based on the U-Net architecture, which consists of a set of down-sampling encoding paths and up-sampling decoding paths. The encoding path consists of multiple 3x3 convolution layers (Conv) and 2x2 max pooling layers (Max Pooling). After each convolution layer, a ReLU (Rectified Linear Unit) activation function is used to increase the non-linear representation 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.

[0058] The decoding path uses a series of 2x2 up-sampling operations to gradually increase the spatial resolution of the feature map, restoring it to its original size. After each up-sampling operation, the feature map is spliced with the corresponding layer in the encoding path to ensure the fusion of global and local features.

[0059] Through skip connections, detailed semantic features from shallow layers can be fused with abstract semantic features from deep layers, allowing the network to balance features of different sizes and depths. Finally, the deep feature map from the last decoding layer is passed through a 3x3 convolution and a Sigmoid activation. The number of channels in the network output can be adjusted according to the segmentation target. Each channel represents a pixel probability map for the corresponding segmentation target, enabling effective fusion of multi-scale features.

[0060] Step 4, the MFP module aims to pre-extract features from input images of different scales. First, the input image is continuously down-sampled to obtain four images of different resolutions. Then, the MFP module is used to extract semantic features from these images. Finally, the obtained feature maps are spliced with the input feature maps of the corresponding encoding layers in the U-Net. Each branch of the MFP module consists of convolution kernels of different sizes and expansion rates (Rate). For example, Figure 2As shown, the first three branches from top to bottom consist of two 3×3 convolutional kernels with dilation rates of 1 and 3, respectively, and a 5×5 convolutional kernel with a dilation rate of 1. The fourth branch connects a standard 3×3 convolutional kernel with a 1×1 convolutional kernel. In the last branch, two cascaded 3×3 convolutional kernels with dilation rates of 3 and 5 are used. As mentioned above, the combination of U-Net and the MFP module allows the network to extract semantic features at different scales in advance, thus enabling the network to detect both large and small targets.

[0061] Step 5: The ESCE path avoids information loss and fully utilizes multi-scale features. For example... Figure 1 As shown, max pooling maps the input feature maps of the second and third encoder layers are mapped to the same feature space as the fourth encoder layer. These feature maps are then concatenated as the input to the ESCE path. The ESCE path consists of three consecutive 3×3 dilated convolution operations with dilation rates of 1, 3, and 5, as shown below. Figure 3 As shown, the feature maps obtained after each convolutional layer are concatenated together to form the output feature map. Finally, the output feature map of the ESCE path is concatenated with the output feature map of the fourth encoder layer to serve as the input to the decoder.

[0062] In step 6, the SE module is built at the decoder input. In the original U-Net model, the encoding and decoding paths are directly connected through skip connections, but there is a semantic gap between them. To mitigate this gap and automatically adjust the contribution of different feature maps, an SE module is inserted at the decoder input. The SE module obtains a 1×1×L one-dimensional feature vector z from the decoder input feature map through global average pooling. Subsequently, z is processed through two fully connected layers (FC layers). The first FC layer has 1×1×(L / 16) neurons and uses ReLU activation function, while the second FC layer has 1×1×L neurons. Finally, a sigmoid activation function is used to fit the complex nonlinear relationship between different channels, generating a 1×1×L channel weighted vector s, which is then multiplied with the original feature map of the corresponding channel to generate the output feature map. Through this method, the decoder path can focus more on salient feature map channels and reduce interference from useless channels.

[0063] In step 7, micro-crack target recognition detection. The data set processed by data augmentation in step 2 is divided into training set and test set according to the proportion of 8:2. In the training process, the cross-entropy loss function is used to measure the difference between the predicted results and the actual labels, and the Adam optimizer is used for model optimization. In order to prevent overfitting, 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 continuous 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. Through the gray image output by the model, the detection result of micro-crack is extracted. In order to further improve the detection accuracy, post-processing technology can be used to refine the detection result, such as removing noise and connecting broken micro-cracks through morphological operation, and using threshold segmentation technology to binarize the gray image to extract the clear micro-crack contour. Finally, the micro-crack detection method based on MES-Net model can generate accurate micro-crack recognition image, which provides important basis for subsequent rock performance analysis and geological disaster warning.

[0064] It can be understood that the present application is described by some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to the features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, the features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

[0065] As, without changing the basic framework of MES-Net, the model performance can be optimized by adjusting the convolution kernel size and expansion rate in MFP module, or using different activation functions or normalization techniques in SE module, so as to achieve similar detection effect.

Claims

1. A deep learning-based rock micro-crack identification and detection method, characterized by: The method comprises the following steps: Step one, the microcracks in the target rock are visualized by vacuum impregnation of fluorescent resin, and a fluorescent image in RGB format is obtained; Step two, the fluorescent image obtained in step one is subjected to data enhancement processing, and a deep learning database is established; Step three, a U-Net network model is established, which comprises a set of down-sampling encoding paths and up-sampling decoding paths, the down-sampling encoding paths comprise a plurality of 3*3 convolution layers Conv and 2*2 maximum pooling layers Max Pooling, and a ReLU linear rectification activation function is used after each convolution layer to increase the nonlinear expression ability of the network; the up-sampling decoding paths gradually increase the spatial resolution of the feature map by a series of 2*2 up-sampling Up-sampling operations, so that the feature map is restored to the original size; Step four, a multi-scale feature pre-extraction MFP module is established for pre-extracting features from input images of different scales to obtain a spliced feature map; Step five, the spliced feature map is taken as the input of an encoder spatial cascade encoding path ESCE, the encoder spatial cascade encoding path ESCE is composed of three consecutive 3*3 dilated convolution operations with dilations of 1, 3 and 5, the feature maps obtained after each layer of convolution are spliced together as the output feature map; the output feature map is spliced with the output feature map of the fourth encoder as the input of the decoder; Step six, a decoder input module SE is inserted into the decoder, and the output feature map is obtained by performing a global average pooling operation, two fully connected layers FCLaye, and a Sigmoid activation function fitting on the channel weighting vector s and multiplying it with the original feature map of the corresponding channel; Step seven, the data sets in the database of step two are divided into a training set and a test set according to a ratio of 8:2, the test data after model training are input into the MES-Net model trained in steps three to six for prediction, and the detection result of the microcracks is extracted from the gray-scale image output by the model.

2. The method according to claim 1, wherein the step of The method for obtaining the spliced feature map is: first, the input image is subjected to continuous down-sampling to obtain four images with different resolutions; then, semantic features are extracted from these images using the MFP module; finally, the obtained feature map is spliced with the input feature map of the corresponding encoding layer in the U-Net.

3. The method of claim 1, wherein the method is characterized by: The model training process in step seven is: a cross-entropy loss function Cross-Entropy Loss is used to measure the difference between the prediction result and the actual label, and an Adam optimizer Adam Optimizer is used for model optimization; an early stopping mechanism EarlyStopping is used to prevent overfitting, and the validation set loss value of each epoch training round is monitored; when the validation set loss value does not significantly decrease in consecutive epochs, the training is stopped in advance.

4. The method of claim 1, wherein the step of identifying and detecting the micro-cracks in the rock based on deep learning is characterized by: The four MFP modules include five branches, the first three branches from top to bottom are composed of two 3*3 convolution kernels with expansion rates of 1 and 3 respectively, and a 5*5 convolution kernel with an expansion rate of 1; the fourth branch connects a 3*3 standard convolution kernel with a 1*1 convolution kernel; the last branch adopts two cascaded 3*3 convolution kernels with expansion rates of 3 and 5 respectively.

Citation Information

Patent Citations

  • Lightweight pavement crack detection method

    CN116503327A

  • Bridge crack detection method based on multi-scale feature fusion and multi-layer attention

    CN118941542A