Device and method for detecting rock disc cracking by using deep learning

Abstract

The present invention provides a rock disc crack detection device, comprising: an image generation unit for extracting boundary information based on Gaussian difference from an original image; an image data input unit that receives the original image and the Gaussian difference image output by the image generation unit; an adversarial generation neural network for generating disc crack data by calculating the image data output by the image data input unit; and a tray crack scene generation unit that generates at least one tray crack scene on the basis of the tray crack data output by the adversarial generation neural network.

Description

Technical Field

[0001] The present invention relates to a rock fracture detection device and method using deep learning, and more particularly, to a formation boundary tracking method and device including a generative adversarial neural network that uses an original image and a Difference of Gaussian (DoG) image as input data. Background Art

[0002] Generally, outdoor rock surveys are conducted by experts directly, which not only consumes a large amount of time and cost, but is also limited in terms of the accessibility and size of the survey area. To solve this problem, deep learning-based technologies for automatically segmenting fractures (or joint traces) from rock surface images are being actively developed.

[0003] In particular, a rock fracture detection method using a Convolutional Neural Network (CNN) (Chen et al., 2021; Byun et al., 2021; Lee et al., 2022). As certified by Chen et al. (2021), compared to existing image processing technologies, a convolutional neural network can segment fractures with higher accuracy and precision. The high performance of a convolutional neural network learns high-level features through convolutional operations within a layer without manually designing features. Thus, a convolutional neural network gradually extracts complex features through a multi-layer structure, thereby providing an output result similar to the manual result.

[0004] Existing rock fracture segmentation networks are mainly based on the U-Net (Ronneberger et al., 2015) architecture, using an encoder and decoder structure to analyze high-level images, classifying the presence of fractures in pixel units, and then restoring the resolution. However, existing networks have problems such as relatively thick expression of fracture lines or noise included in the segmentation results (Byun et al., 2021; Lee et al., 2022). Summary of the Invention Technical Problem

[0005] An object of the present invention is to provide a rock fracture detection device and method that effectively provides accurate results based on an identification loss function and a Difference of Gaussian image using a convolutional neural network model R-DoGAN dedicated to rock fracture detection work. Technical Solution

[0006] The present invention can be implemented in various ways, including a device (system), a method, a computer program stored in a computer-readable medium, or a computer-readable medium storing a computer program.

[0007] The rock mass fracture detection device according to an embodiment of the present invention includes: an image generation unit that extracts boundary information based on the difference of Gaussians from an original image; an image data input unit that receives the original image and the difference-of-Gaussians image output by the image generation unit; a generative adversarial neural network that generates rock mass fracture data by operating on the image data output by the image data input unit; and a rock mass fracture scenario generation unit that generates at least one rock mass fracture scenario based on the rock mass fracture data output by the generative adversarial neural network.

[0008] The image generation unit applies Gaussian blurs of different levels and generates a difference-of-Gaussians image by calculating the differences between the blurred images.

[0009] The generative adversarial neural network includes: a label generation unit that receives the image data output by the image data input unit and extracts rock mass fracture information from the image data to generate a rock mass fracture label; a rock mass fracture discrimination unit that synthesizes a rock mass boundary feature layer by converting the rock mass fracture label output by the label generation unit and the manually recognized rock mass fracture label; and an identification loss function that calculates a permanent loss based on the differences between the rock mass fracture feature layers of the rock mass fracture discrimination unit and re-inputs it to the label generation unit.

[0010] The label generation unit includes a convolutional layer, a downsampling block, a spatial feature aggregation block, a generative residual block, and an upsampling block.

[0011] The rock mass fracture discrimination unit includes a downsampling block and a discrimination residual block.

[0012] The identification loss function is represented by the following formula:

[0013] The generative adversarial neural network includes: a discrimination loss function Loss dis , which receives and processes the image data output by the rock mass fracture discrimination unit; a reconstruction loss function Loss rec , which calculates the loss of the manually recognized rock mass fracture feature map; and a generative loss function Loss gen , which receives the output data of the reconstruction loss function Loss rec to calculate the loss.

[0014] The discrimination loss function Loss dis is a hinge loss function, and the fracture segmentation reconstruction loss function Loss rec is an average hinge loss function.

[0015] A method for detecting rock mass cracks according to an embodiment of the present invention includes the following steps: generating a Difference of Gaussian (DoG) image based on an original image; extracting rock mass crack information using the original image and the DoG image to generate a rock mass crack label; synthesizing a rock mass boundary feature layer by transforming the rock mass crack label and a manually identified rock mass crack label and calculating a permanent loss based on the difference between rock mass crack feature layers; and generating at least one rock mass crack scenario based on the rock mass crack data.

[0016] The step of generating the DoG image includes the following steps: receiving the original image; and generating a DoG image with different intensity difference levels from the original image using a DoG pyramid.

[0017] The step of calculating the permanent loss includes the following steps: synthesizing a rock mass crack feature layer by transforming the rock mass crack label and a manually identified rock mass crack label; and calculating the permanent loss using the manually identified rock mass crack feature map and the rock mass crack feature map. Effects of the Invention

[0018] According to the present invention, more detailed and accurate rock mass crack detection performance can be achieved based on a generative adversarial network using an identification loss function and DoG image data.

[0019] According to the present invention, high-precision detection of rock mass crack detection performance and micro-cracks in the rock mass can be achieved by improving the initial learning efficiency of the generative adversarial network.

[0020] The effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned can be clearly understood by those of ordinary skill in the art to which the present invention pertains (referred to as "ordinary skilled artisans") based on the content recited in the claims of the invention. Brief Description of the Drawings

[0021] Hereinafter, embodiments of the present invention will be described with reference to the following drawings, where like reference numerals denote like structural elements, and are not limited thereto.

[0022] Figure 1 A block diagram showing the structure of a rock mass crack detection device using deep learning according to an embodiment of the present invention.

[0023] Figure 2 A diagram showing the generation result of a DoG image and an original image according to an embodiment of the present invention.

[0024] Figure 3 A diagram showing a generative adversarial network according to an embodiment of the present invention.

[0025] Figure 4 A flowchart showing the operation of a rock mass crack detection method according to an embodiment of the present invention.

[0026] Figure 5 A graph showing the result of comparing the decreasing trend of the loss function during the learning process of the bedrock crack detection device, relativistic generative adversarial network (R-GAN), and baseline model of the present invention.

[0027] Figure 6 A diagram showing the bedrock crack extraction results of the present invention and the prior art.

[0028] Explanation of reference numerals 100: Bedrock crack detection device 110: Image generation unit 120: Image data input unit 130: Generative adversarial neural network 131: Label generation unit 132: Bedrock crack discrimination unit 133: Recognition loss function 134: Discrimination unit loss function 135: Reconstruction loss function 136: Generation unit loss function 140: Bedrock crack scenario generation unit Detailed implementation mode

[0029] Hereinafter, the specific content for implementing the present invention will be described in detail with reference to the accompanying drawings. However, in the following description, when there is a risk of unnecessarily confusing the gist of the present invention, the specific description of known functions or structures will be omitted.

[0030] In the accompanying drawings, the same reference numerals are given to the same or corresponding structural elements. And, during the description of the following embodiments, the repeated description related to the same or corresponding structural elements will be omitted. However, even if the description of the structural elements is omitted, it does not mean that the structural elements are not included in any embodiments.

[0031] In this specification, the advantages, features, and implementation methods of the disclosed embodiments can be made clear by referring to the embodiments described in combination with the accompanying Figure 1 drawings and explanations. However, the present invention is not limited to the embodiments disclosed below, and can be implemented in different embodiments. And, the present embodiments are only provided to enable those of ordinary skill in the technical field to which the present invention pertains to fully understand the scope of the present invention.

[0032] In this specification, unless otherwise defined, all terms (including technical terms and scientific terms) used can be used with the meanings commonly understood by those of ordinary skill in the technical field to which the present invention pertains. And, unless specifically defined, terms generally defined in a dictionary should not be understood as having an idealized or overly formal meaning.

[0033] In this specification, unless clearly specified as singular in the context, singular expressions include plural expressions. Also, unless clearly specified as plural in the context, plural expressions include singular expressions. Throughout the entire content of the specification, when indicating that a certain part includes a certain structural element, unless there is a particularly contrary record, it means that other structural elements are also included, and other structural elements are not excluded.

[0034] In this specification, terms such as "comprising" and "including" indicate the existence of multiple features, multiple steps, multiple operations, multiple elements, and / or structural elements. Such terms do not exclude the additional possibility of one or more other functions, multiple steps, multiple operations, multiple elements, multiple structural elements, and / or their combinations.

[0035] Hereinafter, a rock fracture detection device and method using deep learning according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0036] Figure 1 A block diagram showing the structure of a rock fracture detection device using deep learning according to an embodiment of the present invention. Figure 2 A diagram showing the generation result of a Gaussian difference image and the original image according to an embodiment of the present invention. Figure 3 A diagram showing a generative adversarial network according to an embodiment of the present invention.

[0037] Refer to Figure 1 , a rock fracture detection device 100 using deep learning according to an embodiment of the present invention includes: an image generation unit 110 that extracts boundary information based on Gaussian difference from the original image; an image data input unit 120 that receives the original image and the Gaussian difference image output by the image generation unit; a generative adversarial network 130 that generates rock fracture data by operating on the image data output by the image data input unit; and a rock fracture scene generation unit 140 that generates at least one rock fracture scene based on the rock fracture data output by the generative adversarial network.

[0038] Refer to Figure 2 , after applying Gaussian blurring of different levels, the image generation unit 110 generates a Gaussian difference (DoG, Difference of Gaussian) image by calculating the difference between the blurred images.

[0039] The image of Gaussian difference level 1 refers to the image obtained by removing the first Gaussian standard deviation from the original image, the image of Gaussian difference level 2 refers to the image obtained by removing the second Gaussian standard deviation from the first Gaussian standard deviation, and the image of Gaussian difference level 3 refers to the image obtained by removing the third Gaussian standard deviation from the second Gaussian standard deviation.

[0040] Images from Difference of Gaussian level 1 to Difference of Gaussian level 3 represent boundary information of different intensities.

[0041] Image boundary information is extracted using a Difference of Gaussian (DoG) pyramid. After gradually sampling the images applied with multiple standard Gaussian blurs in the Difference of Gaussian (DoG) pyramid, the differences are calculated to include multi-resolution boundary information. For example, the Gaussian blur steps are set to and generate Difference of Gaussian level images.

[0042] The image data input unit 120 receives the original image and the Difference of Gaussian image output by the image generation unit 110. The image data input unit 120 sets the image size. For example, it receives an image size of 256×256 pixels. When receiving an image larger than 256×256 pixels, the image data input unit 120 outputs an image size of 256×256 pixels by cropping.

[0043] Refer to Figure 3 , the generative adversarial neural network 130 generates the bedrock boundary based on the recognition loss function using the image data output by the image data input unit 120.

[0044] The generative adversarial neural network 130 includes: a label generation unit 131 that receives the image data output by the image data input unit 120 and extracts bedrock crack information from the image data to generate a bedrock crack label; a bedrock crack discrimination unit 132 that synthesizes a bedrock boundary feature layer by converting the bedrock crack label output by the label generation unit 131 and the manually recognized bedrock crack label; and a recognition loss function 133 that calculates the permanent loss based on the difference between the bedrock crack feature layers of the bedrock crack discrimination unit 132 and re-inputs it to the label generation unit 131.

[0045] The label generation unit 131 includes a convolutional layer, a downsampling block, a Spatial Feature Aggregation (SFA) block, a generator residual block, and an upsampling block. The label generation unit 131 marks the pixels corresponding to cracks and the pixels corresponding to the background other than that as 1 and 0 respectively to generate a binary image label of the same size as the bedrock image.

[0046] The bedrock crack discrimination unit 132 synthesizes a bedrock crack feature layer by converting the bedrock crack label output by the label generation unit 131 and the manually recognized bedrock crack label. Thus, the bedrock crack discrimination unit 132 distinguishes the manually recognized bedrock crack feature map from the bedrock crack feature map generated by the label generation unit 131.

[0047] The bedrock crack discrimination unit 132 may include a down sampling block and a discriminator residual block.

[0048] The bedrock crack feature map learns high-dimensional information related to the spatial distribution and morphology of bedrock cracks.

[0049] The recognition loss function (Loss per ) 133 calculates the permanent loss based on the difference before the bedrock crack feature map layer of the bedrock crack discrimination unit 132 and re-inputs it to the label generation unit 131.

[0050] The recognition loss function 133 can be expressed by the following formula.

[0051] Where x is the input image, G(x) is the bedrock crack feature map generated by the label generation unit, y is the manually recognized bedrock crack feature map, and D i is the i-th feature map of the discriminator, and λ i is the weighted value of the bedrock crack feature map.

[0052] For example, λ i can be set to λ1, λ2 = 0.3, λ3, λ4, λ5, λ6 = 0.7.

[0053] Referring to Figure 3 , the generative adversarial neural network 130 may include: a discriminator loss function (Loss dis ) 134, which receives and processes the image data output by the bedrock crack discrimination unit 132; a reconstruction loss function (Loss rec ) 135, which calculates the loss of the manually recognized bedrock crack feature map; and a generator loss function (Loss gen ) 136, which receives the output data of the reconstruction loss function (Loss rec ) 135 to calculate the loss.

[0054] The discriminator loss function (Loss dis ) 134 can be a hinge loss function, and the crack segmentation reconstruction loss function (Loss rec ) 135 of the label generation unit 131 can be a balanced hinge loss function.

[0055] The bedrock crack scene generation unit 140 generates at least one bedrock crack scene based on the bedrock crack data output by the generative adversarial neural network.

[0056] Hereinafter, a method for detecting rock fractures according to an embodiment of the present invention will be described with reference to the accompanying drawings. Figure 4 FIG. is a flowchart showing the operation of a method for detecting rock fractures according to an embodiment of the present invention.

[0057] In the method for detecting rock fractures according to an embodiment of the present invention, after using a generative adversarial neural network to receive an original image and a difference-of-Gaussians (DoG) image to generate synthetic data, the data is used as learning data for rock fractures.

[0058] The method for detecting rock fractures according to an embodiment of the present invention includes a four-step operation process.

[0059] Refer to Figure 4 , the method for detecting rock fractures according to an embodiment of the present invention includes the following steps: generating a difference-of-Gaussians (DoG) image based on the original image; using the original image and the difference-of-Gaussians (DoG) image to extract rock fracture information to generate a rock fracture label; synthesizing a rock boundary feature layer by converting the rock fracture label and the manually identified rock fracture label and calculating a permanent loss based on the difference between the rock fracture feature layers; and generating at least one rock fracture scenario based on the rock fracture data.

[0060] Hereinafter, the process of generating a rock fracture scenario will be described in detail according to the method for detecting rock fractures according to an embodiment of the present invention. First step: Generate a difference of Gaussian image based on the original image (Step S41 0)。

[0061] According to an embodiment of the present invention, the step of generating a difference-of-Gaussians (DoG) image based on the original image includes the following steps: receiving the original image; and using a difference-of-Gaussians (DoG) pyramid to generate a difference-of-Gaussians (DoG) image with different intensity levels from the original image.

[0062] The image of the first difference-of-Gaussians (DoG) level is an image obtained by removing the first Gaussian standard deviation from the original image, the image of the second difference-of-Gaussians (DoG) level is an image obtained by removing the second Gaussian standard deviation from the first Gaussian standard deviation, and the image of the third difference-of-Gaussians (DoG) level is an image obtained by removing the third Gaussian standard deviation from the second Gaussian standard deviation.

[0063] For example, the images of the first to third difference-of-Gaussians (DoG) levels represent boundary information of different intensities.

[0064] The image boundary information is extracted using a difference-of-Gaussians (DoG) pyramid. After gradually sampling the images to which multiple standard Gaussian blurs are applied in the difference-of-Gaussians (DoG) pyramid, the differences are calculated to include multi-resolution boundary information. For example, the Gaussian blur step is set to and a difference-of-Gaussians (DoG) level image is generated. Second step: Generate labels using the original image and the difference of Gaussian image (step S420). forming rock fractures Label (step S420).

[0065] According to an embodiment of the present invention, in the step of generating a bedrock crack label, a convolutional layer, a downsampling block, a spatial feature aggregation (SFA) block, a generator residual block, and an upsampling block are used to label the pixels corresponding to cracks and the pixels corresponding to the background other than the cracks as 1 and 0, respectively, to generate a binary image label having the same size as the bedrock image. Third step: Calculate the permanent loss (step S430).

[0066] According to an embodiment of the present invention, the step of calculating the permanent loss includes the following steps: generating a bedrock crack feature layer by converting the bedrock crack label and the manually identified bedrock crack label; and calculating the permanent loss using the manually identified bedrock crack feature map and the bedrock crack feature map.

[0067] The permanent loss can be calculated through an identification loss function.

[0068] The identification loss function can be expressed by the following formula.

[0069] where x is the input image, G(x) is the bedrock crack feature map generated by the label generator, y is the manually identified bedrock crack feature map, D i is the i-th feature map of the discriminator, and λ i is the weighting value of the bedrock crack feature map. Fourth step: Generate a rock mass fracture scenario (step S440).

[0070] According to an embodiment of the present invention, in the step of generating a bedrock crack scenario, at least one bedrock crack scenario is generated based on the bedrock crack data output by the generative adversarial neural network. Embodiment

[0071] To learn and test the bedrock crack detection apparatus and method according to an embodiment of the present invention, a data set composed of 150 bedrock surface images taken at different locations in Korea (Gwanaksan in Seoul, Bukhansan, Donghae City, Pocheon City, Jeongseon County, etc.) is used. Each image includes a manually marked crack map as ground truth.

[0072] Among the 150 images, 140 are used for learning, and the remaining 10 are used to evaluate the network performance.

[0073] Considering the computational efficiency, each image is segmented into a size of 256×256 pixels.

[0074] Rotation and contrast adjustment were applied for data augmentation, and a total of 11,624 learning images were generated.

[0075] The network was trained to process 8 images at a time (batch size) within 20 epochs.

[0076] Regarding the loss function, as the loss function (Loss dis ) 134 of the rock fracture discrimination unit 132, the hinge loss function was used, and as the fracture segmentation reconstruction loss function (Loss rec ) 135 of the label generation unit 131, the balanced hinge loss function was used.

[0077] Considering the uneven grades, a weighting value inversely proportional to the number of pixels in each grade was applied.

[0078] Furthermore, the recognition loss function Loss per was calculated based on the following formula using the absolute difference between the feature maps of the discriminator.

[0079] The rock fracture device according to the embodiment of the present invention uses both Gaussian difference data and the recognition loss function, while the Baseline model uses a basic Convolutional Neural Network (CNN) model, and the Relativistic Generative Adversarial Network (R-GAN) model only uses the recognition loss function.

[0080] Compared with the Baseline model and the Relativistic Generative Adversarial Network model, the rock fracture detection device (R-DoGAN) 100 according to the embodiment of the present invention has results with relatively less noise and clear fracture segmentation.

[0081] The performance metrics of each model are shown in Table 1 below. Table 1

[0082] Compared with the Baseline model, the Relativistic Generative Adversarial Network model has a higher Intersection-over-Union (IoU) and Precision. The recognition loss function based on the Generative Adversarial Network is beneficial to improving the network performance.

[0083] Furthermore, compared with the Relativistic Generative Adversarial Network, although the present invention has a higher Recall, the Intersection-over-Union (IoU) is slightly lower.

[0084] The Scanline Intersection Similarity (SIS) is an index specifically used to evaluate crack detection work. The SIS of the present invention and the relativistic generative adversarial network model are 0.479 and 0.5 respectively, and the results have a similar level of subjective difference from the manual labels.

[0085] Figure 5 It is a graph showing the results of comparing the decreasing trend of the loss function during the learning process of the rock mass crack detection device, relativistic generative adversarial network (R-GAN), and baseline model of the present invention.

[0086] Refer to Figure 5 , among the three models, the decrease in the reconstruction loss Loss_rec shows a similar trend.

[0087] However, in the case of the recognition loss Loss_per, although there is no obvious change in the baseline model, the rock mass crack device of the present invention and the relativistic generative adversarial network model gradually decrease during the training process.

[0088] This result indicates that the recognition loss function can capture the differences between crack detection results that cannot be solved by the pixel-level loss.

[0089] If the present invention is compared with the relativistic generative adversarial network, the recognition loss of the present invention decreases more rapidly in the initial training step. This is because the input of the difference of Gaussians data improves the initial learning efficiency by providing additional boundary information to the network.

[0090] Figure 6 It is a figure showing the rock mass crack extraction results of the present invention and the prior art.

[0091] Refer to Figure 6 In the right figure of , if the crack segmentation results of the present invention are compared with the existing network (Lee et al., 2022), the results show that the present invention has a higher detection accuracy for thinner crack lines, smaller noises, and micro-cracks.

[0092] The comparison results of the present invention, the relativistic generative adversarial network model, Byun et al. (2021), and Lee et al. (2022) are shown in Tables 2 and 3. Table 2 Table 3

[0093] Compared with Byun et al. (2021), the present invention (R-DoGAN) has a higher recall rate, and compared with Lee et al. (2022), it has better performance in terms of precision.

[0094] All the methods and processes described above can be implemented and fully automated by software code modules executed by one or more general-purpose computers or processors. The code modules can be stored in any type of computer-readable storage medium or other computer storage devices. Part or all of the methods can be implemented by special computer hardware.

[0095] In this specification, any ordinary description, element, or block of the flowcharts illustrated in the description and / or the drawings should be understood as a potential representation of code, module, segment, or part of one or more executable instructions for implementing a specific logical function or element. Instead of examples belonging within the scope of the examples described herein, elements or functions can be deleted, shown, or integrated in a substantially the same or reverse order according to the functions that can be understood herein.

[0096] The embodiments described above can produce various variations and modifications, and such elements should be understood as one of the other permissible examples. All such modifications and variations are included within the scope of the present disclosure and are protected by the scope of the appended claims. The embodiments of the present invention described above can be implemented by program instructions executable by various computer structural elements and can be recorded on a computer-readable recording medium. The computer-readable recording medium can include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the computer-readable recording medium can be instructions specifically designed for the present invention or known instructions that can be used by those of ordinary skill in the computer software field. For example, the computer-readable recording medium can be magnetic media such as hard disks, floppy disks, and magnetic disks, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically designed to store and execute program instructions such as read-only memory, random access memory, and flash memory. For example, the program instructions include not only machine language code generated by a compiler but also high-level language code that can be executed by an interpreter or the like used by a computer. To execute the processing process of the present invention, the hardware device can work through one or more software modules, and vice versa.

[0097] Although the present invention has been described above through specific structural elements, specific embodiments, and drawings, this is only for a further complete understanding of the present invention. The present invention is not limited to the described embodiments, and those of ordinary skill in the technical field to which the present invention pertains can make various technical modifications and variations based on the above description.

[0098] Therefore, the idea of the present invention is not limited to the described embodiments. Except for the scope of the appended claims of the invention, all contents equivalent to the scope of the claims of the invention or equivalent variations fall within the scope of the idea of the present invention.

Claims

1. A rock mass crack detection device, characterized in that, Comprising: An image generation unit that extracts boundary information based on difference of Gaussians from the original image; An image data input unit that receives the original image and the difference-of-Gaussians image output by the image generation unit; A generative adversarial neural network that generates rock mass fracture data by operating on the image data output by the image data input unit; And A rock mass fracture scenario generation unit that generates at least one rock mass fracture scenario based on the rock mass fracture data output by the generative adversarial neural network.

2. The rock mass crack detection device according to claim 1, characterized in that, After applying Gaussian blur of different levels, the image generation unit generates a difference-of-Gaussians image by calculating the differences between the blurred images.

3. The rock mass crack detection device according to claim 1, wherein The generative adversarial neural network includes: A label generation unit that receives the image data output by the image data input unit and extracts rock mass fracture information from the image data to generate a rock mass fracture label; A rock mass fracture discrimination unit that synthesizes a rock mass boundary feature layer by transforming the rock mass fracture label output by the label generation unit and the manually identified rock mass fracture label; and An identification loss function that calculates a permanent loss based on the differences between the rock mass fracture feature layers of the rock mass fracture discrimination unit and re-inputs it to the label generation unit.

4. The rock mass crack detection device according to claim 3, characterized in that, The label generation unit includes a convolutional layer, a downsampling block, a spatial feature aggregation block, a generative residual block, and an upsampling block.

5. The rock mass crack detection device according to claim 3, characterized in that, The rock mass fracture discrimination unit includes a downsampling block and a discrimination residual block.

6. The rock fracture detection device according to claim 3, wherein The identification loss function is represented by the following formula:

7. The rock mass crack detection device according to claim 3, wherein, The generative adversarial neural network includes: Discriminator Loss Function Loss dis , receives and processes the image data output by the rock fracture discriminator; Reconstruction loss function Loss rec , calculate the loss of the manually identified rock mass crack feature map; and Generator Loss Function Loss gen receives the output data of the reconstruction loss function Loss rec to calculate the loss.

8. The rock mass crack detection device according to claim 7, characterized in that, The discrimination part loss function Loss dis is a folding loss function, and the crack segmentation reconstruction loss function Loss rec is an average folding loss function.

9. A method for detecting rock mass cracks, characterized in that, Including the following steps: Generating a difference-of-Gaussians image based on the original image; Using the original image and the difference-of-Gaussians image to extract rock mass fracture information to generate a rock mass fracture label; Synthesizing a rock mass boundary feature layer by transforming the rock mass fracture label and the manually identified rock mass fracture label and calculating a permanent loss based on the differences between the rock mass fracture feature layers; And Generating at least one rock mass fracture scenario based on the rock mass fracture data.

10. The rock mass crack detection method according to claim 9, characterized in that, The step of generating the difference-of-Gaussians image includes the following steps: Receiving the original image; and Using the Gaussian difference pyramid to generate a difference-of-Gaussians image with different intensity difference levels from the original image.

11. The rock mass crack detection method according to claim 9, characterized in that, The step of calculating the permanent loss includes the following steps: Synthesizing a rock mass fracture feature layer by transforming the rock mass fracture label and the manually identified rock mass fracture label; and Calculating the permanent loss using the manually identified rock mass fracture feature map and the rock mass fracture feature map.

Images