Deep neural network segmentation method suitable for rock slice image

Abstract

The invention discloses a deep neural network extraction method suitable for a rock slice image. The method mainly comprises the following steps: adding a light receptive field module behind first five coding layers of a network; according to the network coding layer obtained in the previous step, adding a coordinate attention module behind each light receptive field module; constructing a network decoding layer according to the network coding layer obtained in the previous step to obtain a deep neural network model suitable for rock slice segmentation; using the training data set, and introducing a binary cross entropy loss function to train the deep neural network model constructed in the previous step; and inputting the rock slice image into the deep neural network model trained in the previous step to obtain a segmented image. The rock slice segmentation effect is good, and the method is an effective rock core image particle extraction method.

Description

Technical Field

[0001] The present invention relates to thin slice image segmentation technology, and in particular to a deep neural network extraction method suitable for rock thin slice images, belonging to the field of semantic segmentation. Background Art

[0002] Understanding the porosity, pore size distribution, pore connectivity and other characteristics of rocks is of great significance for the search and exploitation of oil and gas, and the analysis and judgment of these characteristics require the use of rock thin section image segmentation technology. For methods based on image processing technology, they are mainly divided into methods based on traditional image processing technology and methods based on deep learning. Since most images in traditional image segmentation algorithms lack global semantic information, they have poor generalization and cannot cope with complex and changing scenes well. Although the use of traditional image processing technology can greatly improve the efficiency of particle extraction, there are still some shortcomings. Since there are a large number of fine particles in rock thin section images, the edge features between these particles are very similar and cannot be accurately distinguished. At the same time, uneven dyeing during the slicing process will cause the color features of the thin section pores to be unbalanced, resulting in the inability to segment.

[0003] With the development of deep learning image technology, image segmentation methods based on deep learning have far surpassed traditional segmentation methods in terms of segmentation accuracy and speed. The emergence of technologies such as codec basic models, fully convolutional networks, and dilated spatial feature pyramids has made the output image resolution higher after segmentation. For example, Olaf Ronneberger et al. proposed the UNet network structure, which is widely used in the field of medical image segmentation. In 2020, the University of Alberta proposed the UNet based on UNet. 2 Net network structure, which is an RSU structure. Each RSU is a reduced version of UNet. Finally, the complete model is constructed through the jump connection of FPN.

[0004] U 2 Net is a convolutional neural network structure for semantic segmentation. 2 The architecture of Net is a two-layer nested U-structure, which mainly includes residual U-blocks (RSU) for extracting multi-scale features and an outer U-structure connecting RSU, which can capture more contextual information, and the depth of the entire architecture is increased without increasing the computational cost because pooling operations are used in these RSU blocks. 2 The basic processing of UNet network includes input preprocessing, encoder feature extraction, decoder feature recovery, feature fusion, segmentation prediction and post-processing. 2Net can accurately segment the input image into foreground and background objects, and extract rich detail features while retaining context information. 2 Net performs well in processing rock images with complex features. Although RSU and skip connections can extract intra-level and inter-level information of different sizes of images, it is easy to retain invalid features such as non-granular areas and lose information such as defect edges in the connection. Therefore, it is necessary to improve the network and enhance the segmentation ability by improving the ability to extract effective features and reducing the loss of information. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a deep neural network segmentation method suitable for rock thin section images to solve the problems existing in the above-mentioned prior art.

[0006] The present invention proposes a deep neural network extraction method suitable for core particle images, which mainly includes the following steps:

[0007] (1) Add a lightweight receptive field module after the first five encoding layers of the network;

[0008] (2) According to the network coding layer obtained in step (1), a convolutional attention module is embedded after each light receptive field module;

[0009] (3) constructing a network decoding layer according to the network coding layer obtained in step (2) to obtain a deep neural network segmentation model suitable for rock thin section images;

[0010] (4) Using the training data set, a binary cross entropy loss function is introduced to train the deep neural network model constructed in step (3);

[0011] (5) Input the core particle image into the deep neural network model trained in step (4) to obtain the core particle segmentation image. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a lightweight receptive field module.

[0013] Figure 2 It is the convolutional block attention module.

[0014] Figure 3 It is the deep neural network model proposed in the present invention.

[0015] Figure 4 It's U 2 Net network for rock thin section image segmentation.

[0016] Figure 5 This is a diagram showing the effect of the present invention on the segmentation of rock slice images. DETAILED DESCRIPTION

[0017] The present invention will be further described below in conjunction with the accompanying drawings:

[0018] A deep neural network extraction method suitable for core particle images can be divided into the following steps:

[0019] (1) Add a lightweight receptive field module after the first five encoding layers of the network;

[0020] (2) According to the network coding layer obtained in step (1), a convolutional attention module is embedded after each light receptive field module;

[0021] (3) constructing a network decoding layer according to the network coding layer obtained in step (2) to obtain a deep neural network segmentation model suitable for rock thin section images;

[0022] (4) Using the training data set, a binary cross entropy loss function is introduced to train the deep neural network model constructed in step (3);

[0023] (5) Input the rock thin section image into the deep neural network model trained in step (4) to obtain the rock thin section segmentation image.

[0024] Specifically, Figure 1 As shown, in step (1), after the first five feature maps of the network coding layer are output, light receptive field modules with channel numbers of 64, 128, 256 and 512 are added respectively.

[0025] like Figure 2 As shown in Figure 2, in order to suppress the interference of the background area and increase the weight of the target area, a convolutional attention module CA is embedded after each light receptive field module. Given the input image features, CA uses pooling kernels (H, 1) and (1, W) in two different spatial ranges from the vertical and horizontal directions to channel encode the feature map. Therefore, the output of c channels and height h is as shown in Formula 1:

[0026]

[0027] Similarly, the output of the cth channel at width w is as shown in Formula 2:

[0028]

[0029] The above two formulas encode features in the vertical and horizontal directions respectively, aggregate image features, and thus generate a pair of directional attention feature maps. In addition, the above two transformation formulas enable the CA module to effectively capture the long-term dependency in one direction and retain the precise spatial coordinate information in this direction, so that CA can more accurately and effectively calculate the area of ​​interest in the image features.

[0030] Formula 1 and Formula 2 extract global spatial information, increase the receptive field and obtain more accurate spatial coordinate encoding information. After that, CA performs coordinate attention generation to effectively utilize the obtained encoding information. After calculation by Formula 1 and Formula 2, the obtained feature map is input into the shared 1×1 convolution layer, that is, the F1 transformation function, and Formula 3 is obtained:

[0031]

[0032] Among them, [·,·] means that the two parts of feature z h and z w Connect along the spatial dimension, δ is the Sigmoid nonlinear activation function, and the generated is the spatial encoding feature in both vertical and horizontal directions. Among them, r is used to control the block size reduction rate in the SE module. After that, the obtained features are decomposed along the spatial dimension to obtain two independent tensors f h and f w . It is fed into two more 1×1 convolutional layers F h and F w , these two convolutional layers are used to transform f h and f w Converted to features with the same number of channels as the input features, as shown in Formula 4 and Formula 5:

[0033] g h =σ(F h (f h )) (4)

[0034] g w =σ(F w (f w )) (5)

[0035] Among them, δ is still the Sigmoid nonlinear activation function. The adjustment of the reduction rate r can control the number of channels of f, which is conducive to reducing the computational overhead of the model. The output feature g h and g w This is the required attention score. Finally, the output y of the coordinate attention mechanism CA is c It can be expressed as formula 6:

[0036]

[0037] In step (3), the network coding layer obtained in step (2) is used to construct the network decoding layer at the same time, and the network decoding layer is obtained as follows: Figure 3 A deep neural network segmentation model for rock thin section images is shown.

[0038] Since the foreground and background need to be segmented, the binary cross entropy loss function (BCEWithLogitsLoss) is selected as the loss function during model training. The calculation formula of the loss function is as follows:

[0039] BCEWithLogitsLoss=Sigmoid+BCELoss

[0040]

[0041] Among them, Sigmoid represents the activation function, N represents the total number of pixels, and y i represents the value of pixel i in the annotated image, x i Represents the value of pixel i in the predicted image after passing through the model.

[0042] Using the training data set, the binary cross entropy loss function introduced in step (4) is used to train the deep neural network model constructed in step (3).

[0043] In the step (5), the rock thin section image is input into the deep neural network model trained in the step (4) to obtain a segmented image.

[0044] The present invention proposes a deep neural network segmentation method suitable for rock thin section images, which utilizes a lightweight receptive field module to obtain multi-scale information and a convolutional attention module to improve the weight of the target area, and introduces a binary cross entropy loss function to effectively improve the extraction effect of particles in rock thin section images. Figure 4-5 Given U 2 Net method and the effect diagram of the invention on particle segmentation in rock thin section images.

[0045] At the same time, the present invention uses six indicators, including accuracy, intersection over union (IOU), mean pixel accuracy (MPA), mean intersection over union (MIoU), frequency-weighted intersection over union (FWIoU), and F1 score, to evaluate the training effect of the algorithm on the rock thin section dataset. 2 Net method, the experimental results are shown in Table 1. It can be seen that the present invention is better than U 2 The Net method has improved all the segmentation indicators, among which the average intersection-over-union ratio and the frequency-weighted intersection-over-union ratio have been improved significantly, because the invention increases the segmentation capability of the core particles in the image.

[0046] Table 1

[0047] method Accuracy IoU MPA MIoU FWf F1 UNet3+ 93.61 77.74 87.85 82.48 72.88 85.89 The present invention 94.57 80.19 89.08 84.48 75.92 87.56

[0048] In summary, compared with U 2Net method, the present invention has obvious advantages in rock thin slice image segmentation effect and subjective and objective evaluation. Therefore, the present invention is an effective rock thin slice segmentation method.

[0049] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A deep neural network segmentation method suitable for rock thin section images, characterized in that Due to the imbalance of pore color characteristics of rock thin section images, a large number of fine particles, and very similar edge features, a method based on improved U 2 Net network rock thin section image segmentation model, and then use the training data set to train the designed deep neural network model. Finally, the test data set is input into the trained model to obtain a thin section segmentation image with good segmentation effect. The specific steps are as follows: (1) Construct a training set and crop the original core image to obtain a thin slice image with a size of 256×256 pixels; (2) Constructing an improved U 2 Net network, whose input is a thin slice image of size 256×256 pixels; The network improvements are as follows: 2 Net network is a network model based on codec, which is deepened by residual U-shaped modules. First, small receptive field modules (RFB-S) with 64, 128, 256 and 512 channels are added after the first four coding layers of the network. This module first uses three 1×1 convolution kernels to reduce the number of channels of the input feature map, and then uses 1×3, 3×1 and 3×3 convolution kernels to obtain features of different scales. Then, different dilation rates (1, 3 and 5) are used for dilation convolution to expand the receptive field of the network. Finally, the output of the convolution kernel is merged to realize feature fusion. This process can effectively improve the receptive field and feature extraction ability of the network, and can reduce the number of parameters and computational complexity of the network to a certain extent. Then, a coordinate attention module (CA) is embedded after each receptive field module. This module is divided into two parts: Coordinate information embedding and Coordinate Attention generation. The calculation formula is as follows: f=δ(F1([z h ,z w ])) (3) g h =σ(F h (f h )) (4) g w =σ(F w (f w )) (5) in, represents the output of the cth channel with height h, represents the output of the cth channel at width w, and [·,·] represents the combination of the two features z h and z w Connect along the spatial dimension, δ is the Sigmoid nonlinear activation function, and the generated is the spatial encoding feature in both vertical and horizontal directions, r is the reduction rate, and f h and f w The obtained features are decomposed along the spatial dimension to obtain two independent tensors, F h and F w represents two 1×1 convolutional layers, σ is still the Sigmoid nonlinear activation function, g h and g w That is the required attention score. (3) Using the training data set, the binary cross entropy loss function is introduced to train the deep neural network model. The calculation formula of the loss function is as follows: BCEWithLogitsLoss=Sigmoid+BCELoss Among them, Sigmoid represents the activation function, N represents the total number of pixels, and y i represents the value of pixel i in the annotated image, x i Represents the value of pixel i in the predicted image after passing through the model; (4) Input the rock thin section image into the trained deep neural network model to obtain the corresponding segmented image.

Images