Deep neural network extraction method suitable for core particle image

Abstract

The invention discloses a deep neural network extraction method suitable for a core particle image. The method mainly comprises the following steps: adding a receptive field module behind the first four coding layers of a network; according to the network coding layer obtained in the previous step, a convolution attention module is embedded behind each 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 core image particle extraction; 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 core particle image into the deep neural network model trained in the previous step to obtain a particle extraction image in the image. The rock core image particle extraction method is good in rock core image particle extraction effect and is an effective rock core image particle extraction method.

Description

Technical Field

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

[0002] Core particles are an important geological data for studying granular reservoirs. Core particle images usually contain multiple types of mineral particles. The particle segmentation effect is of great significance for geological experts to evaluate relevant geological information. Particle extraction methods mainly include manual annotation methods and methods based on image processing technology. For manual annotation methods, it is mainly completed by manual annotation by geological researchers, which requires a lot of time and effort and is easily affected by subjective factors, resulting in insufficient accuracy of particle extraction. 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. For traditional image processing technology, there are often phenomena such as over-segmentation, under-segmentation, and insufficient smoothness of particle edges when segmenting particles. Although the use of traditional image processing technology can greatly improve the efficiency of particle extraction, there are still some shortcomings. When the core particle image is relatively complex, these methods are difficult to obtain accurate extraction results.

[0003] In recent years, with the rapid development of deep learning, related image processing algorithms have been significantly improved and achieved remarkable results. Long et al. proposed a fully convolutional network FCN, which uses a skip connection structure to make the segmentation results more accurate and refined. Ronneberger et al. proposed a UNet network with a U-shaped structure, which uses skip connections to splice shallow and deep features and has achieved good results in the field of medical image segmentation. Badrinarayanan et al. proposed a SegNet network, which transfers the maximum pooling index to the decoder, further reducing the complexity of segmentation. Zhou et al. proposed a UNet++ network, which adds dense short connections to UNet, making the network deeper and more refined, thereby improving the segmentation effect.

[0004] UNet3+ is a convolutional neural network structure for semantic segmentation, which is an improvement and extension of the classic UNet network. While retaining the basic framework and ideas of UNet, the UNet3+ structure adds more cross-layer information interaction mechanisms to improve the performance and generalization ability of the network. The basic processing process of the UNet3+ network is to use multiple convolutional layers and pooling layers in the encoding process to extract feature representations of different scales to adapt to objects and scenes of different sizes. In the decoding process, cross-layer connections are used to fuse the feature maps extracted in the downsampling process with the feature maps in the upsampling process, so as to better retain the object boundary and local detail information during segmentation, enhance the language information of the deep network, and finally realize the classification of each pixel. However, due to the rich colors of core particle images, large differences in shape and size, and blurred boundaries, it is difficult to extract relatively complete particles by directly using this network to extract core particles, resulting in under-segmentation or over-segmentation of particles. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a deep neural network extraction method suitable for core particle 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 receptive field module after the first four 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 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 extraction model suitable for core particle 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 the receptive field module;

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

[0014] Figure 3 It is the deep neural network model proposed by the present invention;

[0015] Figure 4 This is the effect of UNet3+ network on core image particle segmentation;

[0016] Figure 5 This is a diagram showing the effect of the present invention on the particle segmentation of core 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 receptive field module after the first four 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 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 extraction model suitable for core particle 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 core particle image into the deep neural network model trained in step (4) to obtain the core particle segmentation image.

[0024] Specifically, Figure 1 As shown, in step (1), after the first four feature maps of the network coding layer are output, 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 step (2), in order to suppress the interference of the background area and increase the weight of the target area, a convolutional attention module is embedded after each receptive field module. The following formula is shown:

[0026] M c (F)=σ{MLP[AvgPool(F)]+MLP[MaxPool(F)])}

[0027]

[0028] M s =σ{f 7×7 [AvgPool(F'); MaxPool(F')]}

[0029] Among them, F represents the input feature map, AvgPool and MaxPool represent global average pooling and global maximum pooling respectively, MLP represents the shared fully connected layer, which consists of two fully connected layers and a nonlinear activation function, σ represents the Sigmoid function, Represents the multiplication of corresponding elements of two matrices, f 7×7 Indicates a convolution layer with a convolution kernel size of 7×7.

[0030] 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 extraction model suitable for core particle images is shown.

[0031] 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:

[0032] BCEWithLogitsLoss=Sigmoid+BCELoss

[0033]

[0034] 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.

[0035] 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).

[0036] In the step (5), the core particle image is input into the deep neural network model trained in the step (4) to obtain a particle segmentation image in the image.

[0037] The present invention proposes a deep neural network extraction method suitable for core particle images, which uses a 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 core particle images. Figure 4-5 The UNet3+ method and the effect diagram of the invention on particle extraction in core particle images are given.

[0038] At the same time, the present invention uses three indicators, namely, mean intersection over union (mIoU), mean pixel accuracy (mPA) and frequency weighted intersection over union (FWIoU), to evaluate the training effect of the algorithm in the core particle data set. And compared with the UNet3+ method, the experimental results are shown in Table 1. It can be seen that the present invention has improved all indicators of particle extraction compared with the UNet3+ method. Among them, the improvement of mean intersection over union and frequency weighted intersection over union is more obvious, because the present invention increases the ability to extract core particles in the image.

[0039] Table 1

[0040] method QUR mPA FWf UNet3+ 80.23 91.42 82.02 The present invention 85.66 94.41 87.36

[0041] In summary, compared with the UNet3+ method, the present invention has obvious advantages in the extraction effect of core image particles in both subjective and objective evaluation. Therefore, the present invention is an effective core image particle extraction method.

[0042] 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 extraction method suitable for core particle images, characterized in that Since the core particle images are rich in color, have large differences in shape and size, and have blurred boundaries, a core particle extraction model based on the UNet3+ network is designed, and then the training data set is used to train the designed deep neural network model. Finally, the test data set is input into the trained model to obtain a particle extraction image with better segmentation effect. The specific steps are as follows: (1) Construct a training set and crop the original core image to obtain a particle image with a size of 256×256 pixels; (2) Construct an improved UNet3+ network whose input is a particle image of size 256×256 pixels; The network is improved as follows: the UNet3+ network consists of three parts: encoder, jump connection and decoder. First, receptive field modules (RFBs) with channel numbers of 64, 128, 256 and 512 are added after the first four encoding 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 1×1, 3×3 and 5×5 convolution kernels to extract features from different scales. Then, dilated convolutions with different hole rates (1, 3 and 5) are used to expand the network's receptive field. Finally, the outputs of each convolution layer are integrated to achieve feature fusion. This process can effectively improve the network's receptive field and feature extraction capabilities. Then, a convolutional attention module (CBAM) is embedded after each receptive field module. The module is divided into two parts: channel attention module and spatial attention module. The salient feature map output by the module adaptively learns the correlation between different channels and spaces, and enhances the accuracy of the network in locating particle areas. The calculation formula is as follows: M c (F)=σ{MLP[AvgPool(F)]+MLP[MaxPool(F)])} M s =σ{f 7×7 [AvgPool(F');MaxPool(F')]} Among them, F represents the input feature map, AvgPool and MaxPool represent global average pooling and global maximum pooling respectively, MLP represents the shared fully connected layer, which consists of two fully connected layers and a nonlinear activation function, σ represents the Sigmoid function, Represents the multiplication of corresponding elements of two matrices, f 7×7 Indicates a convolution layer with a convolution kernel size of 7×7; (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 core particle image into the trained deep neural network model to obtain the corresponding particle extraction image.

Images