A method for extracting pores and fractures from core CT images based on deep learning

Abstract

The present invention relates to the field of image processing technology, and to a method for extracting pores and seams from core CT images based on deep learning, which includes the following steps: 1. Data set preparation; 2. Model forward reasoning, passing the CT image into the model to obtain the output of the forward reasoning graph; 3. Loss calculation, the feature graph output by the model and the real data label; 4. Repeating steps 2 and 3, after multiple iterations and judging the real pore and seam categories of the image, and calculating the loss of the image data, optimizing according to the weight of the loss ratio; finally, using the trained model to output the image pore and seam results, thereby realizing the extraction of pores and seams. The overlap degree of the pore and seam image extracted by the method of the present invention and the real image pore and seam on the edge and the accuracy in the middle are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for extracting pores and fractures from core CT images based on deep learning. Background Art

[0002] With the rapid development of my country's economy, the demand for oil and gas resources continues to increase, especially the further deepening of exploration and development. Cores are one of the most important basic geological data in oil and gas exploration and development, providing a basis for decision-making in oil and gas exploration and development. The observation and description of cores play an irreplaceable role in determining lithology, inferring sedimentary environment, and comprehensive research on source-reservoir-cap combination. The analysis and processing of core images has become an important task in core analysis. As a new technology, core digital processing has been widely used. With the help of computer software, the collected core images are analyzed to extract intuitive core data and conduct quantitative core analysis, which has important practical significance for the current oil and gas exploration and development work. Controllable core segmentation is an important preliminary work in core analysis. According to the object information contained in the image, such as grayscale, chromaticity, texture and other features, the image is divided into several non-intersecting areas, so that features with consistency or similarity are distributed in the same area, but show obvious differences between different areas. It is more valuable to calculate the porosity of the core, restore the real core, and make the core with similar pore structure more valuable. Summary of the invention

[0003] The present invention provides a method for extracting pores and fractures from core CT images based on deep learning, which can overcome certain defects of the prior art.

[0004] According to the present invention, a core CT image pore extraction method based on deep learning comprises the following steps:

[0005] 1. Dataset preparation;

[0006] Second, model forward reasoning, passing the CT image into the model to obtain the output of the forward reasoning graph;

[0007] 3. Loss calculation, the feature map output by the model and the real data label;

[0008] 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgment, and optimize the image data according to the loss ratio weight; finally, use the trained model to output the target result, thereby realizing the extraction of holes and seams.

[0009] Preferably, in step one, the core CT is visualized by three-dimensional digitization of the core; random noise is added to the core, and a single picture of the core is obtained by slicing; a standard core image and a corresponding standard original image are obtained by adding pulses and Gaussian white noise; the pore points are marked as 1, and the original image is marked as 0, and the data are grouped by naming to obtain a complete data set; the data set is then sorted according to the ratio of training set: test set: validation set = 8:2:1. At this point, the data preparation stage is completed.

[0010] Preferably, in step 2, the model is performed as follows:

[0011] a) Input image Image gt The format is (N,C,H,W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the DSCA module;

[0012] b) Tier 1 feature acquisition: After DSCA, the output channels are set to 3, 16, 16 respectively, and the tier 1 features are acquired;

[0013] c) Perform maxpooling on the tier1 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 32, 32, 32; finally, obtain the tier2 features;

[0014] d) Perform maxpooling on the tier2 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 64, 64, 64; finally, obtain the tier3 features;

[0015] e) Perform maxpooling on the tier3 features with a convolution kernel of 2*2, and then perform three DSCA operations with output channels of 128, 128, 128; finally, obtain the tier4 features;

[0016] f) Perform maxpooling on the tier4 features with a convolution kernel of 2*2, and then perform three DSCA operations with output channels of 256, 256, 256; finally, obtain the tier5 features;

[0017] g) The tier5 feature is upsampled through deconvolution, the output channel is set to 256, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier4 feature; after three DSCA operations, the output channels are 128, 128, 128; finally, the tier4_1 feature is obtained;

[0018] h) The tier4_1 feature is upsampled through deconvolution, the output channel is set to 128, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier3 feature; after three DSCA operations, the output channels are 64, 64, 64; finally, the tier3_1 feature is obtained;

[0019] i) The tier3_1 feature is upsampled through deconvolution, the output channel is set to 32, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier2 feature; after three DSCA operations, the output channels are 32, 32, 32; finally, the tier2_1 feature is obtained;

[0020] j) The tier2_1 feature is upsampled through deconvolution, the output channels are set to 16, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier1 feature; after three DSCA operations, the output channels are 16, 16, 16; finally, the tier1_1 feature is obtained;

[0021] k) The tier1_1 feature is subjected to three DSCA operations, and the output channels are 16, 1, 1; the final feature map Image is obtained pred .

[0022] Preferably, in step a), the DSCA module calculation steps are as follows:

[0023] 1) The DSCA module consists of Conv+BN+PReLU+Conv+BN+PReLU+CBAM modules, where Conv is a conventional convolution, the kernels of the two convs are 3*3 and 1*1 respectively, and the step size is 1 by default; expansion is a parameter that controls the size of the module, which is used to control the size of the convolution model. After calculation, the sizes are (N, C, H, W), which represent the batch, channel, width, and height of the image respectively;

[0024] 2) The input image first passes through the Conv module, with input channels of C, output channels of C, kernel of 3*3, step size of 1*1, padding of 1 by default, bias of False, and group parameter of C;

[0025] 3) The feature map passes through the BN module with C channels;

[0026] 4) The feature map is activated through the PReLU module;

[0027] 5) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0028] 6) The feature map passes through the BN module, and the number of channels is C*expansion;

[0029] 7) The feature map passes through the CBAM module to calculate the attention features;

[0030] 8) The original features are first passed through the Conv module, with the input channel being C, the output channel being C*expansion, the kernel being 1*1, the step size being 1*1, the padding being 1 by default, and the bias being set to False;

[0031] The final result is obtained by adding the attention features; the shape is (N, C, H, W), thus completing a DSCA operation.

[0032] Preferably, the CBAM module includes two independent submodules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively; the calculation steps of the CBAM module are as follows:

[0033] A) The input feature map F (H×W×C) is subjected to global maxpooling based on width and global average pooling based on height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer;

[0034] B) Send them to a two-layer Conv2d, where the first layer has C input channels, C output channels, 1 kernel, 1 stride, bias = False, and PReLU activation function; the second layer has C input channels, C output channels, 1 kernel, 1 stride, and bias = False. The two layers of the neural network are shared.

[0035] C) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated;

[0036] D) Perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the Spatial attention module;

[0037] E) Use the feature map F output by the Channel attention module as the input feature map of the module; perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps;

[0038] F) Perform concat operation on these two feature maps based on channel;

[0039] G) After a 7×7 convolution operation;

[0040] H) is reduced to 1 channel, i.e. H×W×1; after sigmoid, a spatial attention feature is generated, i.e. M_s;

[0041] I) Multiply the feature by the input feature of the module to obtain the final generated attention feature.

[0042] Preferably, in step 3, the training loss is calculated as follows:

[0043] Calculate the loss cost between the real original image and the high-scoring image calculated by the model, where the loss is calculated as follows:

[0044] ①BCE loss calculation:

[0045]

[0046] l n =-w n [y n log x n +(1-y n )·log(1-x n )]

[0047] represents the total loss, l n is the corresponding loss value of the nth category of the real image; x n Infer the probability value of the category for the model, y n The classification value of the true label. n is the weight coefficient of this class.

[0048] ②Dice Loss calculation:

[0049]

[0050] DiceLoss is a set similarity metric function, usually used to calculate the similarity of two samples (value range [0, 1]), |X∩Y|-the intersection between X and Y; |X| and |Y| represent the number of elements in X and Y respectively. The coefficient 2 in the numerator is because the denominator repeatedly counts the common elements between X and Y. X represents the segmented image, and Y is the segmented image inferred by the model.

[0051] The present invention provides a method for extracting pores and fractures from core CT images based on deep learning. The method uses a codec model to train the CT image to obtain the original image data after the pores and fractures are extracted. The Dice coefficient is used to ensure convergence even in the case of sample imbalance. The pore and fracture images obtained by the method of the present invention are significantly improved compared with the real pore and fracture images in terms of both the smoothness of the boundaries and the overlap of the centers. The use of Dice-Loss is more effective for the problem of class imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a method for extracting pores and fractures from core CT images based on deep learning in Example 1;

[0053] Figure 2 This is a schematic diagram of the model in Example 1;

[0054] Figure 3 This is a schematic diagram of the DSCA module in Example 1;

[0055] Figure 4 This is a schematic diagram of the CBAM module in Example 1. DETAILED DESCRIPTION

[0056] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0057] Example 1

[0058] like Figure 1 As shown, this embodiment provides a method for extracting pores and fractures from core CT images based on deep learning, which includes the following steps:

[0059] 1. Dataset preparation: visualize the core CT through three-dimensional digitization of the core; add random noise to the core, use the slicing method to get a single picture of the core, and obtain the standard CT image and the corresponding standard original image by adding pulses and Gaussian white noise. Mark the hole points as 1 and the original image as 0, and group them by naming to get a complete data set; then divide the data set into training set: test set: validation set = 8:2:1. At this point, the data preparation stage is completed.

[0060] Second, model forward reasoning, passing the CT image into the model to obtain the output of the forward reasoning graph;

[0061] The model proceeds as follows:

[0062] a) The input image Imagegt format is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the DSCA module;

[0063] The calculation steps of the DSCA module are as follows:

[0064] 1) The DSCA module consists of Conv+BN+PReLU+Conv+BN+PreLU+CBAM modules, where conv is a conventional convolution, the kernels of the three convs are 1*1, 3*3, and 1*1, and the step size is 1 by default; the expansion coefficient is set in the module to control the amplification channel multiple of the 3*3 module in DSCA; the sizes are (N, C, H, W), which represent the batch, channel, width, and height of the image respectively;

[0065] 2) The input image first passes through the Conv module, with input channels of C, output channels of C, kernel of 3*3, step size of 1*1, padding of 1 by default, bias of False, and group parameter of C;

[0066] 3) The feature map passes through the BN module with C channels;

[0067] 4) The feature map is activated through the PReLU module;

[0068] 5) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the kernel in the parameter is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False;

[0069] 6) The feature map passes through the BN module, and the number of channels is C*expansion;

[0070] 7) The feature map passes through the CBAM module to calculate the attention features;

[0071] 8) The original features are first passed through the Conv module, with the input channel being C, the output channel being C*expansion, the kernel being 1*1, the step size being 1*1, the padding being 1 by default, and the bias being set to False;

[0072] The final result is obtained by adding the attention features; the shape is (N, C, H, W), thus a DSCA operation is completed.

[0073] The CBAM module contains two independent submodules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively; this not only saves parameters and computing power, but also ensures that it can be integrated into the existing network architecture as a plug-and-play module. The calculation steps of the CBAM module are as follows:

[0074] A) The input feature map F (H×W×C) is subjected to global maxpooling and global average pooling based on width and height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer;

[0075] B) Then send them to a two-layer Conv2d, where the first layer has C input channels, C / r ​​output channels (r is the reduction rate), 1 kernel, 1 stride, bias = False, and PReLU activation function; the second layer has C / r input channels, C output channels, 1 kernel, 1 stride, and bias = False. The two layers of the neural network are shared.

[0076] C) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated;

[0077] D) Finally, perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the SpatialAttention module;

[0078] E) Use the feature map F output by the Channel attention module as the input feature map of this module; first, perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps;

[0079] F) Then perform concat operation (channel splicing) on ​​these two feature maps based on channel;

[0080] G) Then it goes through a 7×7 convolution (7×7 has a larger receptive field than 3×3, and the effect is better) operation;

[0081] H) is reduced to 1 channel, i.e. H×W×1; then sigmoid is used to generate the spatial attention feature, i.e. M_s;

[0082] I) Finally, the feature is multiplied by the input feature of the module to obtain the final generated attention feature.

[0083] b) Acquisition of tier1 features: After DSCA, since DSCA does not change the feature shape, it will not be repeated here. The output channels are set to 3, 16, 16 respectively, and the tier1 features are obtained;

[0084] c) Perform maxpooling on the tier1 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 32, 32, 32; finally, obtain the tier2 features;

[0085] d) Perform maxpooling on the tier2 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 64, 64, 64; finally, obtain the tier3 features;

[0086] e) Perform maxpooling on the tier3 features with a convolution kernel of 2*2, and then perform three DSCA operations with output channels of 128, 128, 128; finally, obtain the tier4 features;

[0087] f) Perform maxpooling on the tier4 features with a convolution kernel of 2*2, and then perform three DSCA operations with output channels of 256, 256, 256; finally, obtain the tier5 features;

[0088] g) The tier5 feature is upsampled through deconvolution, the output channels are set to 256, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier4 feature. After three DSCA operations, the output channels are 128, 128, 128; finally, the tier4_1 feature is obtained;

[0089] h) Upsample the tier4_1 feature through deconvolution, set the output channel to 128, shorten the width and height to twice the previous value, and concatenate it with the copied tier3 feature. After three DSCA operations, the output channels are 64, 64, 64, and finally the tier3_1 feature is obtained;

[0090] i) The tier3_1 feature is upsampled through deconvolution, the output channel is set to 32, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier2 feature. After three DSCA operations, the output channels are 32, 32, 32; finally, the tier2_1 feature is obtained;

[0091] j) The tier2_1 feature is upsampled through deconvolution, the output channels are set to 16, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier1 feature. After three DSCA operations, the output channels are 16, 16, 16; finally, the tier1_1 feature is obtained;

[0092] k) The tier1_1 feature is subjected to three DSCA operations, and the output channels are 16, 1, 1; the final feature map Image is obtained pred .

[0093] 3. Loss calculation: the feature map output by the model and the real data label; the training loss calculation of the generator image data is used as the optimization target, so that the converted high-resolution image is more realistic;

[0094] The training loss is calculated as follows:

[0095] Calculate the loss cost between the real original image and the high-scoring image calculated by the model, where the loss is calculated as follows:

[0096] ①BCE loss calculation:

[0097]

[0098] l n =-w n [y n log x n +(1-y n )·log(1-x n )]

[0099] represents the total loss, l n is the corresponding loss value of the nth category of the real image; x n Infer the probability value of the category for the model, y n The classification value of the true label. n is the weight coefficient of this class.

[0100] ②Dice Loss calculation:

[0101]

[0102] DiceLoss is a set similarity metric function, usually used to calculate the similarity of two samples (value range [0, 1]), |X∩Y|-the intersection between X and Y; |X| and |Y| represent the number of elements in X and Y respectively. The coefficient 2 in the numerator is because the denominator repeatedly counts the common elements between X and Y. X represents the segmented image, and Y is the segmented image inferred by the model.

[0103] 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgment, and optimize the image data according to the loss ratio weight (the experimental ratio is 8:2, that is, the weights are 0.8 and 0.2); finally, use the trained model to output the segmentation results, so as to realize the extraction of holes.

[0104] The present invention provides a method for extracting pores and fractures from core CT images based on deep learning. The method uses a codec model to train the CT image to obtain the original image data after the pores and fractures are extracted. The Dice coefficient is used to ensure convergence even in the case of sample imbalance. The pore and fracture images obtained by the method of the present invention are significantly improved compared with the real pore and fracture images in terms of both the smoothness of the boundaries and the overlap of the centers. The use of Dice-Loss is more effective for the problem of class imbalance.

Claims

1. A method for extracting pores and fractures from core CT images based on deep learning, characterized in that: The following steps are involved:

1. Dataset preparation; 2. Model forward reasoning: Pass the CT image into the model to obtain the output of the forward reasoning graph. The specific steps are as follows: 1) The input image Imagegt format is (N, C, H, W), which represents the batch, channel, width, and height of the image respectively; convolution calculation is performed through the DSCA module, and the specific calculation steps of the DSCA module are as follows: a) The DSCA module consists of Conv+BN+PReLU+Conv+BN+PReLU+CBAM modules, where Conv is a regular convolution, the kernels of the two convs are 3*3 and 1*1 respectively, and the step size is 1 by default; expansion is a parameter that controls the size of the module, which is used to control the size of the convolution model. After calculation, the sizes are (N, C, H, W), which represent the batch, channel, width, and height of the image respectively; b) The input image first passes through the Conv module, with input channels of C, output channels of C, kernel of 3*3, step size of 1*1, padding of 1 by default, bias of False, and group parameter of C; c) The feature map passes through the BN module, and the number of channels is C; d) The feature map passes through the PReLU module for activation; e) The input image first passes through the Conv module, the input channel is C, the output channel is C*expansion, the kernel is 1*1, the step size is 1*1, the padding defaults to 1, and the bias is set to False; f) The feature map passes through the BN module, and the number of channels is C*expansion; d) The feature map passes through the CBAM module to calculate the attention feature; the CBAM module contains two independent sub-modules, namely the channel attention module and the spatial attention module, which perform channel and spatial attention respectively. The specific steps of the CBAM module calculation are as follows: i) The input feature map F (H×W×C) is subjected to global max pooling based on width and global average pooling based on height respectively to obtain two 1×1×C feature maps, and the merged features are sent to the next layer; ii) Send them to a two-layer Conv2d, where the first layer has C channels for input, C / r ​​for output, 1 kernel, 1 stride, bias=False, and ReLU as activation function; the second layer has C / r channels for input, C for output, 1 kernel, 1 stride, and bias=False. The two layers of the neural network are shared. iii) After the sigmoid activation operation, the final channel attention feature, i.e. Mc, is generated; iv) Perform element-wise multiplication on Mc and the input feature map F to generate the input features required by the Spatial attention module; v) Take the feature map F output by the channel attention module as the input feature map of the module; perform a channel-based global max pooling and global average pooling to obtain two H×W×1 feature maps; vi) Perform concat operation on these two feature maps based on channel; vii) After a 7×7 convolution operation; viii) Reduce the dimension to 1 channel, i.e. H×W×1; generate the spatial attention feature, i.e. M_s, through sigmoid, thus completing a CBAM operation; h) Multiply the feature by the input feature of the module to obtain the final generated attention feature; i) The original features are first passed through the Conv module, with the input channel being C, the output channel being C*expansion, the kernel being 1*1, the step size being 1*1, the padding being 1 by default, and the bias being set to False; the final result is obtained by adding the original features to the attention features, with a shape of (N, C, H, W), thus completing a DSCA operation; 2) Tier 1 feature acquisition: After DSCA, the output channels are set to 3, 16, 16 respectively, and the tier 1 features are acquired; 3) Perform maxpooling on the tier1 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 32, 32, 32; finally, obtain the tier2 features; 4) Perform maxpooling on the tier2 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 64, 64, 64; finally, obtain the tier3 features; 5) Perform maxpooling on the tier3 features with a convolution kernel of 2*2, and then perform three DSCA operations with output channels of 128, 128, 128; finally, obtain the tier4 features; 6) Perform maxpooling on the tier4 features with a convolution kernel of 2*2, and then perform three DSCA operations with an output channel of 256, 256, 256; finally, obtain the tier5 features; 7) The tier5 feature is upsampled through deconvolution, the output channel is set to 256, the width and height are shortened to twice the previous ones, and it is spliced ​​with the copied tier4 feature; after three DSCA operations, the output channels are 128, 128, 128; finally, the tier4_1 feature is obtained; 8) The tier4_1 feature is upsampled through deconvolution, the output channel is set to 128, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier3 feature; after three DSCA operations, the output channels are 64, 64, 64; finally, the tier3_1 feature is obtained; 9) The tier3_1 feature is upsampled through deconvolution, the output channel is set to 32, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier2 feature; after three DSCA operations, the output channels are 32, 32, 32; finally, the tier2_1 feature is obtained; 10) The tier2_1 feature is upsampled through deconvolution, the output channel is set to 16, the width and height are shortened to twice the previous ones, and it is concatenated with the copied tier1 feature; after three DSCA operations, the output channels are 16, 16, 16; finally, the tier1_1 feature is obtained; 11) The tier1_1 feature is subjected to three DSCA operations, and the output channels are 16, 1, 1; the final feature map Imagepred is obtained; 3. Loss calculation: obtain the feature map output by the model and the real data label; among them, the loss cost between the real original image and the high-score image calculated by the model is measured, and the loss calculation is as follows: BCE loss calculation: l_n=-w_n[y_n·log(x_n)+(1-y_n)·log(1-x_n)] l(x,y) represents the total loss, l_n is the corresponding loss value of the nth class of the real image; x_n is the probability value of the class inferred by the model, y_n is the classification value of the real label, and w_n is the weight coefficient of the class; Dice Loss calculation: DiceLoss=1-2|X∩Y| / (|X|+|Y|) DiceLoss is a set similarity metric function, usually used to calculate the similarity of two samples (value range is [0,1]), |X∩Y| is the intersection between X and Y, |X| and |Y| represent the number of elements in X and Y respectively; the coefficient 2 in the numerator is because the denominator repeatedly calculates the common elements between X and Y, X represents the segmented image, and Y is the segmented image inferred by the model; 4. Repeat steps 2 and 3, after multiple iterations and image resolution category judgments, and optimize the image data according to the loss ratio weight, and finally use the trained optimal model to output the target hole and seam results, thereby realizing the extraction of holes and seams.

2. The method for extracting pores and fractures from core CT images based on deep learning according to claim 1, characterized in that: The step 1 uses a core three-dimensional digital mode to visualize the core CT image, uses a slicing method to obtain a single core picture, obtains a standard image and a corresponding standard original image, marks the hole and fracture points as 1, marks the original image as 0, and groups them through different naming methods to obtain a complete data set; then the data set is divided into training set: test set: validation set = 8:2:1, and the data preparation stage is completed.

Images